{ "query": { "task": "LLM generation" }, "task": "LLM generation", "size_class": "LLM ≤4B", "batch": "1", "mixed": false, "mixed_warning": null, "groups": [ { "size": "LLM ≤4B", "batch": "1", "platforms": 64, "rows": 914 }, { "size": "LLM 5–14B", "batch": "1", "platforms": 53, "rows": 668 }, { "size": "LLM 15–40B", "batch": "1", "platforms": 22, "rows": 82 }, { "size": "LLM >40B", "batch": "1", "platforms": 11, "rows": 18 }, { "size": "LLM 5–14B", "batch": "8", "platforms": 4, "rows": 27 }, { "size": "LLM ≤4B", "batch": "8", "platforms": 4, "rows": 31 }, { "size": "LLM 15–40B", "batch": "8", "platforms": 3, "rows": 29 }, { "size": "LLM >40B", "batch": "8", "platforms": 1, "rows": 3 }, { "size": "LLM >40B", "batch": "unknown", "platforms": 4, "rows": 5 }, { "size": "LLM 15–40B", "batch": "unknown", "platforms": 1, "rows": 1 }, { "size": "unknown", "batch": "1", "platforms": 8, "rows": 577 }, { "size": "unknown", "batch": "8", "platforms": 4, "rows": 15 } ], "notes": [], "count": 64, "results": [ { "platform": "Intel 265K Custom PC (Nvidia RTX 4090) (as tested by Jeff Geerling)", "platform_url": "/platform/intel-265k-custom-pc-nvidia-rtx-4090-as-tested-by-jeff-geerling.html", "perf": 334.67, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 261.9, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 1.278, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/43", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/43", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.8746, "unknown_badges": [ "price unknown" ] }, { "platform": "Intel 13900K (Nvidia 5090) (as tested by Jeff Geerling)", "platform_url": "/platform/intel-13900k-nvidia-5090-as-tested-by-jeff-geerling.html", "perf": 271.4, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "D", "confidence_label": "D – vendor-published", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/pull/18", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.8109, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "Intel Core Ultra 7 368H", "platform_url": "/platform/intel-core-ultra-7-368h.html", "perf": 250, "perf_unit": "tok/s", "metric": "2nd-token throughput, input 32 tokens", "model": "minicpm4-0.5b", "precision": "INT4-MIXED", "stack": "OpenVINO 2026.4.1 GenAI", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "D", "confidence_label": "D – vendor-published", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "AI PC", "vendor": "Intel", "source": { "name": "OpenVINO official LLM benchmarks (llm_models_7-368H.csv)", "url": "https://github.com/openvinotoolkit/openvino/tree/2026.4.1/docs/sphinx_setup/_static/benchmarks_files", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.747, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "Ryzen 9 7900X (Nvidia 4090) (as tested by Jeff Geerling)", "platform_url": "/platform/ryzen-9-7900x-nvidia-4090-as-tested-by-jeff-geerling.html", "perf": 237.05, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "D", "confidence_label": "D – vendor-published", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/pull/11", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.7083, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "Intel 13900K (Nvidia 4090) (as tested by Jeff Geerling)", "platform_url": "/platform/intel-13900k-nvidia-4090-as-tested-by-jeff-geerling.html", "perf": 216.48, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "D", "confidence_label": "D – vendor-published", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/pull/11", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.6468, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "Ryzen 9 7950X (Nvidia 4080) (as tested by Jeff Geerling)", "platform_url": "/platform/ryzen-9-7950x-nvidia-4080-as-tested-by-jeff-geerling.html", "perf": 204.45, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "D", "confidence_label": "D – vendor-published", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/pull/11", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.6109, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "Ryzen 9 7950X (Nvidia 4070 Ti Super) (as tested by Jeff Geerling)", "platform_url": "/platform/ryzen-9-7950x-nvidia-4070-ti-super-as-tested-by-jeff-geerling.html", "perf": 198.95, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "D", "confidence_label": "D – vendor-published", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/pull/11", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.5945, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "Pi CM5 - 16GB (Nvidia RTX 4090) (as tested by Jeff Geerling)", "platform_url": "/platform/pi-cm5-16gb-nvidia-rtx-4090-as-tested-by-jeff-geerling.html", "perf": 235.61, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 290.2, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.8119, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/43", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/43", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.59, "unknown_badges": [ "price unknown" ] }, { "platform": "Intel 265K Custom PC (Nvidia RTX 3080 Ti) (as tested by Jeff Geerling)", "platform_url": "/platform/intel-265k-custom-pc-nvidia-rtx-3080-ti-as-tested-by-jeff-geerling.html", "perf": 263.24, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 495, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.5318, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/39", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/39", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.5491, "unknown_badges": [ "price unknown" ] }, { "platform": "Dell XPS 13 (Intel Core 5 320) (as tested by Jeff Geerling)", "platform_url": "/platform/dell-xps-13-intel-core-5-320-as-tested-by-jeff-geerling.html", "perf": 23.88, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 14, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 1.706, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/52", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/52", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.5357, "unknown_badges": [ "price unknown" ] }, { "platform": "Intel 265K Custom PC (Nvidia RTX 4070 Ti) (as tested by Jeff Geerling)", "platform_url": "/platform/intel-265k-custom-pc-nvidia-rtx-4070-ti-as-tested-by-jeff-geerling.html", "perf": 196.15, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 266.9, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.7349, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/45", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/45", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.5084, "unknown_badges": [ "price unknown" ] }, { "platform": "Ryzen 9 5950X (Nvidia 4070) (as tested by Jeff Geerling)", "platform_url": "/platform/ryzen-9-5950x-nvidia-4070-as-tested-by-jeff-geerling.html", "perf": 160.72, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "D", "confidence_label": "D – vendor-published", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/pull/11", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.4802, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "Intel 265K Custom PC (Nvidia RTX 2080 Ti) (as tested by Jeff Geerling)", "platform_url": "/platform/intel-265k-custom-pc-nvidia-rtx-2080-ti-as-tested-by-jeff-geerling.html", "perf": 180.38, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 257.8, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.6997, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/46", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/46", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.4746, "unknown_badges": [ "price unknown" ] }, { "platform": "Pi CM5 - 16GB (Nvidia RTX 3080 Ti) (as tested by Jeff Geerling)", "platform_url": "/platform/pi-cm5-16gb-nvidia-rtx-3080-ti-as-tested-by-jeff-geerling.html", "perf": 208.5, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 394, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.5292, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/39", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/39", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.4666, "unknown_badges": [ "price unknown" ] }, { "platform": "Intel 265K Custom PC (AMD Radeon AI Pro R9700) (as tested by Jeff Geerling)", "platform_url": "/platform/intel-265k-custom-pc-amd-radeon-ai-pro-r9700-as-tested-by-jeff-geerling.html", "perf": 195.91, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 333.8, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.5869, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/35", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/35", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.4647, "unknown_badges": [ "price unknown" ] }, { "platform": "Apple M4 (10-core GPU, 120 GB/s)", "platform_url": "/platform/apple-m4-10-core-gpu-120-gb-s.html", "perf": 41.31, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 30.1, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 1.372, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/2", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/2", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.4638, "unknown_badges": [ "price unknown" ] }, { "platform": "Pi CM5 - 16GB (Nvidia RTX 4070 Ti) (as tested by Jeff Geerling)", "platform_url": "/platform/pi-cm5-16gb-nvidia-rtx-4070-ti-as-tested-by-jeff-geerling.html", "perf": 166.06, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 228, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.7283, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/45", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/45", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.4615, "unknown_badges": [ "price unknown" ] }, { "platform": "NVIDIA Jetson AGX Thor", "platform_url": "/platform/nvidia-jetson-agx-thor.html", "perf": 122, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Qwen3-VL-2B", "precision": "NVFP4", "stack": "vLLM", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 130, "power_label": "platform spec envelope (not measured with this result)", "power_kind": "spec", "perf_per_w": 0.9385, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "max TDP (OpenNav); price mypcrig", "price_usd": 3499, "price_label": "indicative list price (third-party/vendor)", "form_factor": "edge module", "vendor": "NVIDIA", "source": { "name": "NVIDIA Jetson AI Lab benchmarks (benchmarks.json)", "url": "https://github.com/NVIDIA-AI-IOT/jetson-ai-lab/blob/main/src/data/benchmarks.json", "date": null }, "power_source": "https://mypcrig.com/blog/jetson-orin-thor-local-llm-benchmark-2026/", "price_source": "https://mypcrig.com/blog/jetson-orin-thor-local-llm-benchmark-2026/", "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.4573, "unknown_badges": [] }, { "platform": "Pi CM5 - 16GB (Nvidia RTX A4000) (as tested by Jeff Geerling)", "platform_url": "/platform/pi-cm5-16gb-nvidia-rtx-a4000-as-tested-by-jeff-geerling.html", "perf": 134.47, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 162.7, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.8265, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/36", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/36", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.4431, "unknown_badges": [ "price unknown" ] }, { "platform": "Mac Studio (M3 Ultra 512GB) (as tested by Jeff Geerling)", "platform_url": "/platform/mac-studio-m3-ultra-512gb-as-tested-by-jeff-geerling.html", "perf": 154.6, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 223, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.6933, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/42", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/42", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.4342, "unknown_badges": [ "price unknown" ] }, { "platform": "Intel Core 5 330", "platform_url": "/platform/intel-core-5-330.html", "perf": 136.9863014, "perf_unit": "tok/s", "metric": "2nd-token throughput, input 32 tokens", "model": "minicpm4-0.5b", "precision": "INT4-MIXED", "stack": "OpenVINO 2026.4.1 GenAI", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "D", "confidence_label": "D – vendor-published", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "AI PC", "vendor": "Intel", "source": { "name": "OpenVINO official LLM benchmarks (llm_models_5-330.csv)", "url": "https://github.com/openvinotoolkit/openvino/tree/2026.4.1/docs/sphinx_setup/_static/benchmarks_files", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.4093, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "Pi CM5 - 16GB (Nvidia RTX 2080 Ti) (as tested by Jeff Geerling)", "platform_url": "/platform/pi-cm5-16gb-nvidia-rtx-2080-ti-as-tested-by-jeff-geerling.html", "perf": 155.06, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 258.2, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.6005, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/46", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/46", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.4077, "unknown_badges": [ "price unknown" ] }, { "platform": "Intel Core Ultra 7 258V (Lunar Lake)", "platform_url": "/platform/intel-core-ultra-7-258v-lunar-lake.html", "perf": 135.1351351, "perf_unit": "tok/s", "metric": "2nd-token throughput, input 32 tokens", "model": "minicpm4-0.5b", "precision": "INT4-MIXED", "stack": "OpenVINO 2026.4.1 GenAI", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "D", "confidence_label": "D – vendor-published", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "AI PC", "vendor": "Intel", "source": { "name": "OpenVINO official LLM benchmarks (llm_models_7-258V.csv)", "url": "https://github.com/openvinotoolkit/openvino/tree/2026.4.1/docs/sphinx_setup/_static/benchmarks_files", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.4038, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "Ryzen 9 9950X (AMD 7900 XT) (as tested by Jeff Geerling)", "platform_url": "/platform/ryzen-9-9950x-amd-7900-xt-as-tested-by-jeff-geerling.html", "perf": 131.2, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "D", "confidence_label": "D – vendor-published", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/pull/11", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.392, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "Intel 265K Custom PC (Nvidia RTX A4000) (as tested by Jeff Geerling)", "platform_url": "/platform/intel-265k-custom-pc-nvidia-rtx-a4000-as-tested-by-jeff-geerling.html", "perf": 149.44, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 269.2, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.5551, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/36", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/36", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.386, "unknown_badges": [ "price unknown" ] }, { "platform": "M1 Ultra (64 GPU Core) 128GB (as tested by Jeff Geerling)", "platform_url": "/platform/m1-ultra-64-gpu-core-128gb-as-tested-by-jeff-geerling.html", "perf": 127.6, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/48", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.3813, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "Dell Pro Max with GB10", "platform_url": "/platform/dell-pro-max-with-gb10.html", "perf": 97.93, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 128.5, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.7621, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/34", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/34", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.3697, "unknown_badges": [ "price unknown" ] }, { "platform": "NVIDIA Jetson AGX Orin 64GB", "platform_url": "/platform/nvidia-jetson-agx-orin-64gb.html", "perf": 60, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Qwen3-VL-2B", "precision": "W4A16", "stack": "vLLM", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 65, "power_label": "platform spec envelope (not measured with this result)", "power_kind": "spec", "perf_per_w": 0.9231, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "max TDP per OpenNav report / mypcrig price", "price_usd": 1999, "price_label": "indicative list price (third-party/vendor)", "form_factor": "edge module", "vendor": "NVIDIA", "source": { "name": "NVIDIA Jetson AI Lab benchmarks (benchmarks.json)", "url": "https://github.com/NVIDIA-AI-IOT/jetson-ai-lab/blob/main/src/data/benchmarks.json", "date": null }, "power_source": "https://opennav.org/news/opennav-robotics-workload-benchmark/", "price_source": "https://opennav.org/news/opennav-robotics-workload-benchmark/", "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.3602, "unknown_badges": [] }, { "platform": "Intel 265K Custom PC (Nvidia RTX 3060) (as tested by Jeff Geerling)", "platform_url": "/platform/intel-265k-custom-pc-nvidia-rtx-3060-as-tested-by-jeff-geerling.html", "perf": 122.85, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 214, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.5741, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/40", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/40", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.3518, "unknown_badges": [ "price unknown" ] }, { "platform": "Pi CM5 - 16GB (Nvidia RTX 3060) (as tested by Jeff Geerling)", "platform_url": "/platform/pi-cm5-16gb-nvidia-rtx-3060-as-tested-by-jeff-geerling.html", "perf": 112.77, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 192.3, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.5864, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/40", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/40", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.3403, "unknown_badges": [ "price unknown" ] }, { "platform": "Intel Core Ultra 7 255H", "platform_url": "/platform/intel-core-ultra-7-255h.html", "perf": 109.8901099, "perf_unit": "tok/s", "metric": "2nd-token throughput, input 32 tokens", "model": "minicpm4-0.5b", "precision": "INT4-MIXED", "stack": "OpenVINO 2026.4.1 GenAI", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "D", "confidence_label": "D – vendor-published", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "AI PC", "vendor": "Intel", "source": { "name": "OpenVINO official LLM benchmarks (llm_models_7-255H.csv)", "url": "https://github.com/openvinotoolkit/openvino/tree/2026.4.1/docs/sphinx_setup/_static/benchmarks_files", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.3284, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "Minisforum MS-R1 (Nvidia RTX A2000) (as tested by Jeff Geerling)", "platform_url": "/platform/minisforum-ms-r1-nvidia-rtx-a2000-as-tested-by-jeff-geerling.html", "perf": 71.36, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 94.3, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.7567, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/32", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/32", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.3284, "unknown_badges": [ "price unknown" ] }, { "platform": "AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)", "platform_url": "/platform/amd-ryzen-ai-max-395-pro-395-strix-halo-radeon-8060s.html", "perf": 88.14, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 133, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.6627, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/21#issuecomment-3164568218", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/21#issuecomment-3164568218", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.3259, "unknown_badges": [ "price unknown" ] }, { "platform": "M1 Ultra (48 GPU Core) 64GB (as tested by Jeff Geerling)", "platform_url": "/platform/m1-ultra-48-gpu-core-64gb-as-tested-by-jeff-geerling.html", "perf": 108.67, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "D", "confidence_label": "D – vendor-published", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/pull/11", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.3247, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "NVIDIA Jetson Thor T4000", "platform_url": "/platform/nvidia-jetson-thor-t4000.html", "perf": 95, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Qwen3-VL-2B", "precision": "NVFP4", "stack": "vLLM", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "edge module", "vendor": "NVIDIA", "source": { "name": "NVIDIA Jetson AI Lab benchmarks (benchmarks.json)", "url": "https://github.com/NVIDIA-AI-IOT/jetson-ai-lab/blob/main/src/data/benchmarks.json", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.2839, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "Pi 500+ - 16GB (AMD RX 7900 XT) (as tested by Jeff Geerling)", "platform_url": "/platform/pi-500-16gb-amd-rx-7900-xt-as-tested-by-jeff-geerling.html", "perf": 108.58, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 315, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.3447, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/23", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/23", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.2632, "unknown_badges": [ "price unknown" ] }, { "platform": "System76 Thelio Astra (Nvidia A4000) (as tested by Jeff Geerling)", "platform_url": "/platform/system76-thelio-astra-nvidia-a4000-as-tested-by-jeff-geerling.html", "perf": 90.92, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 244, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.3726, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/5", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/5", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.245, "unknown_badges": [ "price unknown" ] }, { "platform": "System76 Thelio Astra (AMD Pro W7700) (as tested by Jeff Geerling)", "platform_url": "/platform/system76-thelio-astra-amd-pro-w7700-as-tested-by-jeff-geerling.html", "perf": 89.31, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 261, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.3422, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/5", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/5", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.2337, "unknown_badges": [ "price unknown" ] }, { "platform": "Pi 5 - 8GB (AMD RX 6700 XT) 12GB (as tested by Jeff Geerling)", "platform_url": "/platform/pi-5-8gb-amd-rx-6700-xt-12gb-as-tested-by-jeff-geerling.html", "perf": 49.01, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 94, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.5214, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/1", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/1", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.226, "unknown_badges": [ "price unknown" ] }, { "platform": "Pi 500+ - 16GB (AMD RX 9070 XT) (as tested by Jeff Geerling)", "platform_url": "/platform/pi-500-16gb-amd-rx-9070-xt-as-tested-by-jeff-geerling.html", "perf": 89.63, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 304, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.2948, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/25", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/25", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.2203, "unknown_badges": [ "price unknown" ] }, { "platform": "Pi 5 - 8GB (AMD Pro W7700) (as tested by Jeff Geerling)", "platform_url": "/platform/pi-5-8gb-amd-pro-w7700-as-tested-by-jeff-geerling.html", "perf": 56.14, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 145, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.3872, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/9", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/9", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.1974, "unknown_badges": [ "price unknown" ] }, { "platform": "Pi CM5 - 16GB (AMD AI Pro R9700) (as tested by Jeff Geerling)", "platform_url": "/platform/pi-cm5-16gb-amd-ai-pro-r9700-as-tested-by-jeff-geerling.html", "perf": 77.54, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 281, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.2759, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/35", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/35", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.1967, "unknown_badges": [ "price unknown" ] }, { "platform": "Pi 5 - 8GB (AMD RX 6500 XT) (as tested by Jeff Geerling)", "platform_url": "/platform/pi-5-8gb-amd-rx-6500-xt-as-tested-by-jeff-geerling.html", "perf": 39.82, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 88, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.4525, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/1", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/1", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.1921, "unknown_badges": [ "price unknown" ] }, { "platform": "M1 Max Mac Studio (10 core - 64GB) (as tested by Jeff Geerling)", "platform_url": "/platform/m1-max-mac-studio-10-core-64gb-as-tested-by-jeff-geerling.html", "perf": 59.38, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "D", "confidence_label": "D – vendor-published", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks#findings", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.1774, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "Framework 13 (Ryzen AI 5 340) (as tested by Jeff Geerling)", "platform_url": "/platform/framework-13-ryzen-ai-5-340-as-tested-by-jeff-geerling.html", "perf": 23.81, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 51.1, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.4659, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/33", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/33", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.1721, "unknown_badges": [ "price unknown" ] }, { "platform": "Radxa Dragon Q8B (as tested by Jeff Geerling)", "platform_url": "/platform/radxa-dragon-q8b-as-tested-by-jeff-geerling.html", "perf": 12.3, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 24.3, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.5062, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/54", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/54", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.1667, "unknown_badges": [ "price unknown" ] }, { "platform": "Pi 500+ - 16GB (Intel Arc B580) (as tested by Jeff Geerling)", "platform_url": "/platform/pi-500-16gb-intel-arc-b580-as-tested-by-jeff-geerling.html", "perf": 47.38, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 146, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.3245, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/26", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/26", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.1659, "unknown_badges": [ "price unknown" ] }, { "platform": "Pi 5 - 8GB (AMD RX 7600) (as tested by Jeff Geerling)", "platform_url": "/platform/pi-5-8gb-amd-rx-7600-as-tested-by-jeff-geerling.html", "perf": 48.47, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 156, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.3107, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/1", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/1", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.1635, "unknown_badges": [ "price unknown" ] }, { "platform": "Pi 500+ - 16GB (Intel Arc Pro B50) (as tested by Jeff Geerling)", "platform_url": "/platform/pi-500-16gb-intel-arc-pro-b50-as-tested-by-jeff-geerling.html", "perf": 29.8, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 78.5, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.3796, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/27", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/27", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.1558, "unknown_badges": [ "price unknown" ] }, { "platform": "MetaComputing AI PC Framework 13 Mainboard (as tested by Jeff Geerling)", "platform_url": "/platform/metacomputing-ai-pc-framework-13-mainboard-as-tested-by-jeff-geerling.html", "perf": 13.1, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 29, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.4517, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/51", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/51", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.152, "unknown_badges": [ "price unknown" ] }, { "platform": "Pi 500+ - 16GB (as tested by Jeff Geerling)", "platform_url": "/platform/pi-500-16gb-as-tested-by-jeff-geerling.html", "perf": 5.55, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 13, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.4269, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/24", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/24", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.1334, "unknown_badges": [ "price unknown" ] }, { "platform": "Raspberry Pi 5 (CPU only)", "platform_url": "/platform/raspberry-pi-5-cpu-only.html", "perf": 4.88, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 11.9, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.4101, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/70", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/70", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.1275, "unknown_badges": [ "price unknown" ] }, { "platform": "HiFive Premier P550 (AMD RX 580) (as tested by Jeff Geerling)", "platform_url": "/platform/hifive-premier-p550-amd-rx-580-as-tested-by-jeff-geerling.html", "perf": 36.23, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 150, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.2415, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/20", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/20", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.1249, "unknown_badges": [ "price unknown" ] }, { "platform": "Minisforum MS-R1 (as tested by Jeff Geerling)", "platform_url": "/platform/minisforum-ms-r1-as-tested-by-jeff-geerling.html", "perf": 12.12, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 35, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.3463, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/32", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/32", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.1196, "unknown_badges": [ "price unknown" ] }, { "platform": "System76 Thelio Astra (Nvidia A400) (as tested by Jeff Geerling)", "platform_url": "/platform/system76-thelio-astra-nvidia-a400-as-tested-by-jeff-geerling.html", "perf": 35.51, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 167, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.2126, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/5", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/5", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.1154, "unknown_badges": [ "price unknown" ] }, { "platform": "GMKtec G3 Plus (Intel N150) - 16GB (as tested by Jeff Geerling)", "platform_url": "/platform/gmktec-g3-plus-intel-n150-16gb-as-tested-by-jeff-geerling.html", "perf": 9.06, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 26.4, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.3432, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/12", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/12", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.1141, "unknown_badges": [ "price unknown" ] }, { "platform": "Pi 500+ - 16GB (Intel Arc A310 ECO) (as tested by Jeff Geerling)", "platform_url": "/platform/pi-500-16gb-intel-arc-a310-eco-as-tested-by-jeff-geerling.html", "perf": 13.36, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "llama.cpp (Vulkan)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 50, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.2672, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/29", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/29", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.0983, "unknown_badges": [ "price unknown" ] }, { "platform": "Pi 400 - 4GB (as tested by Jeff Geerling)", "platform_url": "/platform/pi-400-4gb-as-tested-by-jeff-geerling.html", "perf": 1.6, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 6, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.2667, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/commit/96bab78f2a8e6c996c6810c5e2119274e3eb401a", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/commit/96bab78f2a8e6c996c6810c5e2119274e3eb401a", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.0806, "unknown_badges": [ "price unknown" ] }, { "platform": "NVIDIA Jetson Orin NX 16GB", "platform_url": "/platform/nvidia-jetson-orin-nx-16gb.html", "perf": 25, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Cosmos Reasoning 2 2B", "precision": "W4A16", "stack": "vLLM", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "edge module", "vendor": "NVIDIA", "source": { "name": "NVIDIA Jetson AI Lab benchmarks (benchmarks.json)", "url": "https://github.com/NVIDIA-AI-IOT/jetson-ai-lab/blob/main/src/data/benchmarks.json", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.0747, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "AmpereOne A192-32X (512GB) (as tested by Jeff Geerling)", "platform_url": "/platform/ampereone-a192-32x-512gb-as-tested-by-jeff-geerling.html", "perf": 23.52, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/10", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.0703, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "NVIDIA Jetson Orin Nano 8GB", "platform_url": "/platform/nvidia-jetson-orin-nano-8gb.html", "perf": 18, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Nemotron3 Nano 4B", "precision": "Q4_K_M GGUF", "stack": "llama.cpp", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": null, "power_label": "no power published", "power_kind": "unknown", "perf_per_w": null, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "unknown", "price_usd": null, "price_label": "unknown", "form_factor": "edge module", "vendor": "NVIDIA", "source": { "name": "NVIDIA Jetson AI Lab benchmarks (benchmarks.json)", "url": "https://github.com/NVIDIA-AI-IOT/jetson-ai-lab/blob/main/src/data/benchmarks.json", "date": null }, "power_source": null, "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.0538, "unknown_badges": [ "power unknown", "perf/W unknown", "price unknown" ] }, { "platform": "HiFive Premier P550 (4-core RISC-V) (as tested by Jeff Geerling)", "platform_url": "/platform/hifive-premier-p550-4-core-risc-v-as-tested-by-jeff-geerling.html", "perf": 0.24, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 13.5, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.01778, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/17", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/17", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.0056, "unknown_badges": [ "price unknown" ] }, { "platform": "Dell Optiplex 780 (C2Q Q8400) (as tested by Jeff Geerling)", "platform_url": "/platform/dell-optiplex-780-c2q-q8400-as-tested-by-jeff-geerling.html", "perf": 1.09, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 146, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.007466, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/31", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/31", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.0038, "unknown_badges": [ "price unknown" ] }, { "platform": "DC-ROMA Mainboard II (8-core RISC-V) (as tested by Jeff Geerling)", "platform_url": "/platform/dc-roma-mainboard-ii-8-core-risc-v-as-tested-by-jeff-geerling.html", "perf": 0.31, "perf_unit": "tok/s", "metric": "generation_throughput", "model": "Llama 3.2:3b", "precision": "as in linked issue (Ollama default tags are 4-bit)", "stack": "Ollama / llama.cpp (see linked issue)", "size_class": "LLM ≤4B", "batch": "1", "like_for_like": true, "differs": [], "confidence": "C", "confidence_label": "C – community/review", "power_w": 30.6, "power_label": "reported in result source — measured", "power_kind": "measured", "perf_per_w": 0.01013, "date": "undated", "older_than_18_months": false, "superseded": false, "power_basis": "measured peak system power (Jeff Geerling, as published)", "price_usd": null, "price_label": "unknown", "form_factor": "independent test system", "vendor": "(independent test)", "source": { "name": "Jeff Geerling ai-benchmarks README (Findings tables)", "url": "https://github.com/geerlingguy/ai-benchmarks/issues/28", "date": null }, "power_source": "https://github.com/geerlingguy/ai-benchmarks/issues/28", "price_source": null, "fit": { "perf": null, "power": null, "budget": null }, "meets_all": true, "unknowns": 0, "eligible": true, "headroom": null, "rank_score": 0.0034, "unknown_badges": [ "price unknown" ] } ], "no_data": null, "legend": "Confidence: A audited MLPerf closed, B documented third-party, C community, D vendor-published, E estimate. Unknown is never assumed." }