Independent tests

312 results from 11 independent testers: people and papers that tested the hardware themselves, not the vendor. Each links to the original.

Jeff Geerling (183)

Tester's site · Independent tester. Eval rate (token generation) as published; details in linked issue/PR.

Results from Jeff Geerling
Platform	Workload	Model	Result	Power	Date	Original
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	DeepSeek R1 14b	11.37 tok/s	140 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Llama 3.2:3b	88.14 tok/s	133 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Llama 3.1:70b	4.97 tok/s	133 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
AmpereOne A192-32X (192 core - 512GB) (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	23.52 tok/s	not reported	Older data (2024)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
AmpereOne A192-32X (192 core - 512GB) (as tested by Jeff Geerling)	LLM generation	llama3.1:8b	17.47 tok/s	not reported	Older data (2024)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
AmpereOne A192-32X (192 core - 512GB) (as tested by Jeff Geerling)	LLM generation	llama3.1:70b	3.86 tok/s	not reported	Older data (2024)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
AmpereOne A192-32X (192 core - 512GB) (as tested by Jeff Geerling)	LLM generation	llama3.1:405b	0.900 tok/s	not reported	Older data (2024)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
AmpereOne A192-32X (512GB) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 671b	4.18 tok/s	477 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
AmpereOne A192-32X (512GB) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	23.52 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
AmpereOne A192-32X (512GB) (as tested by Jeff Geerling)	LLM generation	Llama 3.1:70b	3.86 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
AmpereOne A192-32X (512GB) (as tested by Jeff Geerling)	LLM generation	Llama 3.1:405b	0.900 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Apple M4 (10-core GPU, 120 GB/s)	LLM generation	Llama 3.2:3b	41.31 tok/s	30.10 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
DC-ROMA Mainboard II (8-core RISC-V) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	0.310 tok/s	30.60 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
DC-ROMA Mainboard II (8-core RISC-V) (as tested by Jeff Geerling)	LLM generation	deepseek-r1:7b	4.90 tok/s	38.90 W	2025-10-10 (page last edited)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
DC-ROMA RISC-V Mainboard II (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	0.310 tok/s	not reported	2025-10-17 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
DC-ROMA RISC-V Mainboard II (as tested by Jeff Geerling)	LLM generation	deepseek-r1:1.5b	0.590 tok/s	not reported	2025-10-17 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
DC-ROMA RISC-V Mainboard II (as tested by Jeff Geerling)	LLM generation	deepseek-r1:7b	2.53 tok/s	not reported	2025-10-24 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Dell Optiplex 780 (C2Q Q8400) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	1.09 tok/s	146 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Dell Optiplex 780 (Intel Core 2 Quad Q8400) (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	1.09 tok/s	not reported	2025-11-05 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Dell Optiplex 780 (Intel Core 2 Quad Q8400) (as tested by Jeff Geerling)	LLM generation	tinyllama:1.1b	4.04 tok/s	not reported	2025-11-05 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Dell Optiplex 780 (Intel Core 2 Quad Q8400) (as tested by Jeff Geerling)	LLM generation	deepseek-r1:1.5b	2.30 tok/s	not reported	2025-11-05 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Dell Pro Max with GB10	LLM generation	DeepSeek R1 14b	23.92 tok/s	138 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Dell Pro Max with GB10	LLM generation	gpt-oss 20b	87.97 tok/s	123 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Dell Pro Max with GB10	LLM generation	Llama 3.2:3b	97.93 tok/s	129 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Dell Pro Max with GB10	LLM generation	Llama 3.1:70b	4.71 tok/s	156 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Dell XPS 13 (Intel Core 5 320) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	23.88 tok/s	14.00 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Framework 13 (Ryzen AI 5 340 16GB) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	5.83 tok/s	50.40 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Framework 13 (Ryzen AI 5 340) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	23.81 tok/s	51.10 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Framework 13 (Ryzen AI 5 340) (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	23.81 tok/s	51.10 W	2025-11-12 (page last edited)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
Framework 13 (Ryzen AI 5 340) (as tested by Jeff Geerling)	LLM generation	llama3.1:8b	10.46 tok/s	51.70 W	2025-11-12 (page last edited)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
Framework 13 (Ryzen AI 5 340) (as tested by Jeff Geerling)	LLM generation	deepseek-r1:14b	5.83 tok/s	50.40 W	2025-11-12 (page last edited)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
Framework Desktop Mainboard and 4-node cluster (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	45.36 tok/s	152 W	2025-08-07 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Framework Desktop Mainboard and 4-node cluster (as tested by Jeff Geerling)	LLM generation	llama3.1:8b	20.89 tok/s	142 W	2025-08-07 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Framework Desktop Mainboard and 4-node cluster (as tested by Jeff Geerling)	LLM generation	llama2:13b	12.65 tok/s	142 W	2025-08-07 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Framework Desktop Mainboard and 4-node cluster (as tested by Jeff Geerling)	LLM generation	deepseek-r1:1.5b	85.38 tok/s	142 W	2025-08-07 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Framework Desktop Mainboard and 4-node cluster (as tested by Jeff Geerling)	LLM generation	deepseek-r1:8b	19.49 tok/s	141 W	2025-08-07 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Framework Desktop Mainboard and 4-node cluster (as tested by Jeff Geerling)	LLM generation	deepseek-r1:14b	11.37 tok/s	140 W	2025-08-07 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Framework Desktop Mainboard and 4-node cluster (as tested by Jeff Geerling)	LLM generation	deepseek-r1:70b	2.43 tok/s	141 W	2025-08-07 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Framework Mainboard Cluster (512GB) (as tested by Jeff Geerling)	LLM generation	Llama 3.1:405b	0.710 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
GMKtec G3 Plus (Intel N150) - 16GB (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	9.06 tok/s	26.40 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
GMKtek G3 Plus (Intel N150) - 16GB (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	2.13 tok/s	30.30 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
HiFive Premier P550 (4-core RISC-V) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	0.240 tok/s	13.50 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
HiFive Premier P550 (AMD RX 580) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	36.23 tok/s	150 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
HiFive Premier P550 (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	0.240 tok/s	13.50 W	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
HiFive Premier P550 (as tested by Jeff Geerling)	LLM generation	llama3.1:8b	0.100 tok/s	13.60 W	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
HiFive Premier P550 (as tested by Jeff Geerling)	LLM generation	llama2:13b	0.110 tok/s	13.60 W	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
HiFive Premier P550 (as tested by Jeff Geerling)	LLM generation	deepseek-r1:1.5b	0.490 tok/s	13.50 W	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
Intel 13900K (Nvidia 4090) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	216 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 13900K (Nvidia 5090) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	271 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (AMD Radeon AI Pro R9700) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	53.72 tok/s	331 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (AMD Radeon AI Pro R9700) (as tested by Jeff Geerling)	LLM generation	gpt-oss 20b	163 tok/s	333 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (AMD Radeon AI Pro R9700) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	196 tok/s	334 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (Nvidia RTX 2080 Ti) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	45.06 tok/s	308 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (Nvidia RTX 2080 Ti) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	180 tok/s	258 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (Nvidia RTX 3060) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	29.77 tok/s	224 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (Nvidia RTX 3060) (as tested by Jeff Geerling)	LLM generation	gpt-oss 20b	83.96 tok/s	229 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (Nvidia RTX 3060) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	123 tok/s	214 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (Nvidia RTX 3080 Ti) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	70.74 tok/s	495 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (Nvidia RTX 3080 Ti) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	263 tok/s	495 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (Nvidia RTX 4070 Ti) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	49.74 tok/s	311 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (Nvidia RTX 4070 Ti) (as tested by Jeff Geerling)	LLM generation	gpt-oss 20b	163 tok/s	299 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (Nvidia RTX 4070 Ti) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	196 tok/s	267 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (Nvidia RTX 4090) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	92.49 tok/s	464 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (Nvidia RTX 4090) (as tested by Jeff Geerling)	LLM generation	gpt-oss 20b	280 tok/s	509 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (Nvidia RTX 4090) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	335 tok/s	262 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (Nvidia RTX A4000) (as tested by Jeff Geerling)	LLM generation	gpt-oss 20b	93.55 tok/s	199 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel 265K Custom PC (Nvidia RTX A4000) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	149 tok/s	269 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Intel N150 Mini PC (GMKTek G3 Plus) (as tested by Jeff Geerling)	LLM generation	deepseek-r1:1.5b	17.02 tok/s	not reported	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
Intel N150 Mini PC (GMKTek G3 Plus) (as tested by Jeff Geerling)	LLM generation	deepseek-r1:8b	3.55 tok/s	not reported	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
Intel N150 Mini PC (GMKTek G3 Plus) (as tested by Jeff Geerling)	LLM generation	deepseek-r1:14b	1.97 tok/s	not reported	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
Intel N150 Mini PC (GMKTek G3 Plus) (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	8.04 tok/s	not reported	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
Intel N150 Mini PC (GMKTek G3 Plus) (as tested by Jeff Geerling)	LLM generation	llama3.1:8b	3.59 tok/s	not reported	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
Intel N150 Mini PC (GMKTek G3 Plus) (as tested by Jeff Geerling)	LLM generation	llama2:13b	2.16 tok/s	not reported	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
Intel N150 Mini PC (GMKTek G3 Plus) (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	9.06 tok/s	26.40 W	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
Intel N150 Mini PC (GMKTek G3 Plus) (as tested by Jeff Geerling)	LLM generation	llama3.1:8b	3.91 tok/s	29.80 W	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
Intel N150 Mini PC (GMKTek G3 Plus) (as tested by Jeff Geerling)	LLM generation	llama2:13b	2.57 tok/s	28.50 W	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
Intel N150 Mini PC (GMKTek G3 Plus) (as tested by Jeff Geerling)	LLM generation	deepseek-r1:1.5b	17.99 tok/s	29.90 W	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
Intel N150 Mini PC (GMKTek G3 Plus) (as tested by Jeff Geerling)	LLM generation	deepseek-r1:8b	3.84 tok/s	29.80 W	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
Intel N150 Mini PC (GMKTek G3 Plus) (as tested by Jeff Geerling)	LLM generation	deepseek-r1:14b	2.13 tok/s	30.30 W	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
M1 Max Mac Studio (10 core - 64GB) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	59.38 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
M1 Max Mac Studio (10 core - 64GB) (as tested by Jeff Geerling)	LLM generation	Llama 3.1:70b	7.25 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
M1 Ultra (48 GPU Core) 64GB (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	35.89 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
M1 Ultra (48 GPU Core) 64GB (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	109 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
M1 Ultra (64 GPU Core) 128GB (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	41.51 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
M1 Ultra (64 GPU Core) 128GB (as tested by Jeff Geerling)	LLM generation	gpt-oss 20b	100 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
M1 Ultra (64 GPU Core) 128GB (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	128 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
M1 Ultra (64 GPU Core) 128GB (as tested by Jeff Geerling)	LLM generation	Llama 3.1:70b	9.84 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
M4 Mac mini (10 core CPU) / 32GB (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	41.31 tok/s	30.10 W	Older data (2024)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
M4 Mac mini (10 core CPU) / 32GB (as tested by Jeff Geerling)	LLM generation	llama3.1:8b	20.95 tok/s	29.40 W	Older data (2024)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
M4 Mac mini (10 core CPU) / 32GB (as tested by Jeff Geerling)	LLM generation	llama2:13b	13.60 tok/s	29.80 W	Older data (2024)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
Mac Studio (M3 Ultra 512GB) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	51.85 tok/s	227 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Mac Studio (M3 Ultra 512GB) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 671b	19.89 tok/s	262 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Mac Studio (M3 Ultra 512GB) (as tested by Jeff Geerling)	LLM generation	gpt-oss 20b	115 tok/s	227 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Mac Studio (M3 Ultra 512GB) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	155 tok/s	223 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Mac Studio (M3 Ultra 512GB) (as tested by Jeff Geerling)	LLM generation	Llama 3.1:70b	14.08 tok/s	243 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Mac Studio (M3 Ultra 512GB) (as tested by Jeff Geerling)	LLM generation	Llama 3.1:405b	1.96 tok/s	256 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
MetaComputing AI PC Framework 13 Mainboard (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	3.25 tok/s	32.00 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
MetaComputing AI PC Framework 13 Mainboard (as tested by Jeff Geerling)	LLM generation	gpt-oss 20b	8.52 tok/s	27.00 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
MetaComputing AI PC Framework 13 Mainboard (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	13.10 tok/s	29.00 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Minisforum MS-R1 (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	3.39 tok/s	38.40 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Minisforum MS-R1 (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	12.12 tok/s	35.00 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Minisforum MS-R1 (as tested by Jeff Geerling)	LLM generation	Llama 3.1:70b	0.770 tok/s	38.20 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Minisforum MS-R1 (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	12.12 tok/s	not reported	2025-11-10 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Minisforum MS-R1 (as tested by Jeff Geerling)	LLM generation	llama3.1:8b	5.78 tok/s	not reported	2025-11-10 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Minisforum MS-R1 (as tested by Jeff Geerling)	LLM generation	llama2:13b	3.91 tok/s	not reported	2025-11-10 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Minisforum MS-R1 (as tested by Jeff Geerling)	LLM generation	llama3.1:70b	0.770 tok/s	not reported	2025-11-10 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Minisforum MS-R1 (as tested by Jeff Geerling)	LLM generation	deepseek-r1:14b	3.39 tok/s	not reported	2025-11-10 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Minisforum MS-R1 (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	71.36 tok/s	not reported	2025-11-10 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Minisforum MS-R1 (as tested by Jeff Geerling)	LLM generation	llama3.1:8b	34.52 tok/s	not reported	2025-11-10 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Minisforum MS-R1 (Nvidia RTX A2000) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	71.36 tok/s	94.30 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Nvidia A400 (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	35.51 tok/s	167 W	Older data (2024)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
Nvidia A400 (as tested by Jeff Geerling)	LLM generation	llama3.2:8b	2.79 tok/s	190 W	Older data (2024)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
Nvidia A400 (as tested by Jeff Geerling)	LLM generation	llama2:13b	7.93 tok/s	223 W	Older data (2024)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
Pi 400 - 4GB (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	1.60 tok/s	6.00 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi 5 - 16GB (AMD Pro W7700) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	19.90 tok/s	164 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi 5 - 8GB (AMD Pro W7700) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	56.14 tok/s	145 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi 5 - 8GB (AMD RX 6500 XT) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	39.82 tok/s	88.00 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi 5 - 8GB (AMD RX 6700 XT) 12GB (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	49.01 tok/s	94.00 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi 5 - 8GB (AMD RX 7600) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	48.47 tok/s	156 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi 500 - 8GB (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	4.53 tok/s	10.40 W	Older data (2024)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
Pi 500 - 8GB (as tested by Jeff Geerling)	LLM generation	llama3.2:8b	1.90 tok/s	10.60 W	Older data (2024)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
Pi 500+ - 16GB (AMD RX 7900 XT) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	109 tok/s	315 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi 500+ - 16GB (AMD RX 9070 XT) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	89.63 tok/s	304 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi 500+ - 16GB (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	5.55 tok/s	13.00 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi 500+ - 16GB (Intel Arc A310 ECO) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	13.36 tok/s	50.00 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi 500+ - 16GB (Intel Arc B580) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	47.38 tok/s	146 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi 500+ - 16GB (Intel Arc Pro B50) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	29.80 tok/s	78.50 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (AMD AI Pro R9700) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	7.98 tok/s	280 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (AMD AI Pro R9700) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	77.54 tok/s	281 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (AMD Radeon AI Pro R9700) (as tested by Jeff Geerling)	LLM generation	gpt-oss 20b	19.59 tok/s	318 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (Nvidia RTX 2080 Ti) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	41.23 tok/s	275 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (Nvidia RTX 2080 Ti) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	155 tok/s	258 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (Nvidia RTX 3060) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	29.40 tok/s	194 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (Nvidia RTX 3060) (as tested by Jeff Geerling)	LLM generation	gpt-oss 20b	79.14 tok/s	181 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (Nvidia RTX 3060) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	113 tok/s	192 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (Nvidia RTX 3080 Ti) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	67.74 tok/s	476 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (Nvidia RTX 3080 Ti) (as tested by Jeff Geerling)	LLM generation	gpt-oss 20b	144 tok/s	434 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (Nvidia RTX 3080 Ti) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	209 tok/s	394 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (Nvidia RTX 4070 Ti) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	48.24 tok/s	253 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (Nvidia RTX 4070 Ti) (as tested by Jeff Geerling)	LLM generation	gpt-oss 20b	133 tok/s	235 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (Nvidia RTX 4070 Ti) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	166 tok/s	228 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (Nvidia RTX 4090) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	83.31 tok/s	398 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (Nvidia RTX 4090) (as tested by Jeff Geerling)	LLM generation	gpt-oss 20b	189 tok/s	322 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (Nvidia RTX 4090) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	236 tok/s	290 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (Nvidia RTX A4000) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	36.21 tok/s	163 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (Nvidia RTX A4000) (as tested by Jeff Geerling)	LLM generation	gpt-oss 20b	90.68 tok/s	163 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Pi CM5 - 16GB (Nvidia RTX A4000) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	134 tok/s	163 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Radxa Dragon Q8B (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	12.30 tok/s	24.30 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Radxa Orion O6 - 16GB (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	4.33 tok/s	34.70 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Radxa Orion O6 - 16GB (Nvidia RTX 3080 Ti) (as tested by Jeff Geerling)	LLM generation	DeepSeek R1 14b	64.58 tok/s	465 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Radxa Oryon O6 Mini ITX Board (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	9.33 tok/s	30.30 W	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
Radxa Oryon O6 Mini ITX Board (as tested by Jeff Geerling)	LLM generation	llama3.1:8b	4.95 tok/s	32.10 W	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
Radxa Oryon O6 Mini ITX Board (as tested by Jeff Geerling)	LLM generation	deepseek-r1:1.5b	14.04 tok/s	29.80 W	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
Radxa Oryon O6 Mini ITX Board (as tested by Jeff Geerling)	LLM generation	deepseek-r1:8b	4.87 tok/s	31.80 W	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
Radxa Oryon O6 Mini ITX Board (as tested by Jeff Geerling)	LLM generation	deepseek-r1:14b	2.63 tok/s	31.80 W	Older data (2025)	Jeff Geerling ai-benchmarks issue results (comments)
Radxa Oryon O6 Mini ITX Board (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	16.49 tok/s	34.00 W	2025-04-27 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Radxa Oryon O6 Mini ITX Board (as tested by Jeff Geerling)	LLM generation	llama3.1:8b	7.60 tok/s	34.20 W	2025-04-27 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Radxa Oryon O6 Mini ITX Board (as tested by Jeff Geerling)	LLM generation	llama2:13b	4.17 tok/s	32.80 W	2025-04-27 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Radxa Oryon O6 Mini ITX Board (as tested by Jeff Geerling)	LLM generation	deepseek-r1:1.5b	31.98 tok/s	30.90 W	2025-04-27 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Radxa Oryon O6 Mini ITX Board (as tested by Jeff Geerling)	LLM generation	deepseek-r1:8b	7.90 tok/s	37.30 W	2025-04-27 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Radxa Oryon O6 Mini ITX Board (as tested by Jeff Geerling)	LLM generation	deepseek-r1:14b	4.33 tok/s	34.70 W	2025-04-27 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Radxa Oryon O6 Mini ITX Board (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	181 tok/s	409 W	2025-05-06 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Radxa Oryon O6 Mini ITX Board (as tested by Jeff Geerling)	LLM generation	llama3.1:8b	113 tok/s	471 W	2025-05-06 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Radxa Oryon O6 Mini ITX Board (as tested by Jeff Geerling)	LLM generation	llama2:13b	84.21 tok/s	453 W	2025-05-06 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Radxa Oryon O6 Mini ITX Board (as tested by Jeff Geerling)	LLM generation	deepseek-r1:1.5b	214 tok/s	320 W	2025-05-06 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Radxa Oryon O6 Mini ITX Board (as tested by Jeff Geerling)	LLM generation	deepseek-r1:8b	112 tok/s	466 W	2025-05-06 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Radxa Oryon O6 Mini ITX Board (as tested by Jeff Geerling)	LLM generation	deepseek-r1:14b	64.58 tok/s	465 W	2025-05-06 (page last edited)	Jeff Geerling ai-benchmarks issue results (comments)
Raspberry Pi 400 - 4GB (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	1.60 tok/s	6.00 W	Older data (2024)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
Raspberry Pi 5 (CPU only)	LLM generation	DeepSeek R1 14b	1.20 tok/s	13.00 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Raspberry Pi 5 (CPU only)	LLM generation	Llama 3.2:3b	4.88 tok/s	11.90 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Raspberry Pi 5 (CPU only)	LLM generation	Llama 3.2:3b	4.61 tok/s	13.90 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Raspberry Pi CM5 - 4GB (as tested by Jeff Geerling)	LLM generation	llama3.2:3b	4.58 tok/s	not reported	Older data (2024)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
Raspberry Pi CM5 - 8GB (as tested by Jeff Geerling)	LLM generation	llama3.1:8b	1.93 tok/s	not reported	Older data (2024)	Jeff Geerling sbc-reviews (AI / LLM inference tables)
Raspberry Pi CM5 Cluster (10x 16GB) (as tested by Jeff Geerling)	LLM generation	Llama 3.1:70b	0.850 tok/s	70.00 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Ryzen 9 5950X (Nvidia 4070) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	161 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Ryzen 9 7900X (Nvidia 4090) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	237 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Ryzen 9 7900X (Nvidia 4090) (as tested by Jeff Geerling)	LLM generation	Llama 3.1:70b	3.10 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Ryzen 9 7950X (Nvidia 4070 Ti Super) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	199 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Ryzen 9 7950X (Nvidia 4080) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	204 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
Ryzen 9 9950X (AMD 7900 XT) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	131 tok/s	not reported	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
System76 Thelio Astra (AMD Pro W7700) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	89.31 tok/s	261 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
System76 Thelio Astra (Nvidia A400) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	35.51 tok/s	167 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)
System76 Thelio Astra (Nvidia A4000) (as tested by Jeff Geerling)	LLM generation	Llama 3.2:3b	90.92 tok/s	244 W	2026-09-22 (page last edited)	Jeff Geerling ai-benchmarks README (Findings tables)

arXiv 2507.04842 authors (32)

Tester's site ·

Results from arXiv 2507.04842 authors
Platform	Workload	Model	Result	Power	Date	Original
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n (Hard-Swish, xView3-SAR)	6.90 FPS	5.40 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n (Hard-Swish, xView3-SAR)	24.70 FPS	15.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n (Hard-Swish, xView3-SAR)	12.80 FPS	6.40 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n (Hard-Swish, xView3-SAR)	29.80 FPS	15.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n (Hard-Swish, xView3-SAR)	19.80 FPS	7.60 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n (Hard-Swish, xView3-SAR)	29.80 FPS	15.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n (Hard-Swish, xView3-SAR)	21.60 FPS	7.80 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n (Hard-Swish, xView3-SAR)	29.80 FPS	15.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n-Ghost (Hard-Swish, xView3-SAR)	7.20 FPS	5.10 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n-Ghost (Hard-Swish, xView3-SAR)	27.00 FPS	15.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n-Ghost (Hard-Swish, xView3-SAR)	13.40 FPS	6.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n-Ghost (Hard-Swish, xView3-SAR)	33.60 FPS	15.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n-Ghost (Hard-Swish, xView3-SAR)	20.90 FPS	7.20 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n-Ghost (Hard-Swish, xView3-SAR)	33.60 FPS	15.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n-Ghost (Hard-Swish, xView3-SAR)	22.90 FPS	7.30 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n-Ghost (Hard-Swish, xView3-SAR)	33.60 FPS	15.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n-Ghost-P2 (Hard-Swish, xView3-SAR)	3.60 FPS	4.90 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n-Ghost-P2 (Hard-Swish, xView3-SAR)	13.30 FPS	15.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n-Ghost-P2 (Hard-Swish, xView3-SAR)	7.00 FPS	5.70 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n-Ghost-P2 (Hard-Swish, xView3-SAR)	20.50 FPS	15.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n-Ghost-P2 (Hard-Swish, xView3-SAR)	11.80 FPS	6.90 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n-Ghost-P2 (Hard-Swish, xView3-SAR)	20.50 FPS	15.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n-Ghost-P2 (Hard-Swish, xView3-SAR)	13.40 FPS	7.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8n-Ghost-P2 (Hard-Swish, xView3-SAR)	20.50 FPS	15.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8s-Ghost-P2 (Hard-Swish, xView3-SAR)	3.20 FPS	5.30 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8s-Ghost-P2 (Hard-Swish, xView3-SAR)	9.30 FPS	15.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8s-Ghost-P2 (Hard-Swish, xView3-SAR)	6.20 FPS	6.50 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8s-Ghost-P2 (Hard-Swish, xView3-SAR)	12.20 FPS	15.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8s-Ghost-P2 (Hard-Swish, xView3-SAR)	10.40 FPS	8.10 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8s-Ghost-P2 (Hard-Swish, xView3-SAR)	12.20 FPS	15.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8s-Ghost-P2 (Hard-Swish, xView3-SAR)	10.50 FPS	8.10 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6
AMD Kria KV260 Vision AI Starter Kit (K26 SOM)	object detection	YOLOv8s-Ghost-P2 (Hard-Swish, xView3-SAR)	12.20 FPS	15.00 W	2025-11-17 (published)	arXiv 2507.04842v2 (SAR vessel detection on Kria KV260), Table 6

visorcraft (GitHub) (28)

Tester's site · llama.cpp llama-bench pp512 / tg128 (rows marked * measured via llama-server); Fedora 43, 85W/120W power profiles tested (power not measured per row)

Results from visorcraft (GitHub)
Platform	Workload	Model	Result	Power	Date	Original
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	TinyLlama 1.1B	249 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Llama 3.2 3B	60.90 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Llama 2 7B	47.30 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Qwen2.5-Coder 7B	36.70 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Qwen2.5 14B	24.50 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Qwen3 30B-A3B MoE	86.10 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Qwen2.5 32B	11.30 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Llama 3.1 70B	5.10 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Qwen3-Coder-Next 80B-A3B	42.70 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	GPT-OSS 120B	53.40 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	MiniMax M2.1-REAP-139B	29.30 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	MiniMax M2.5	32.80 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Qwen3-235B-A22B	17.20 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Nemotron-3 Nano 30B-A3B	61.50 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	TinyLlama 1.1B	6513 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Llama 3.2 3B	2248 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Llama 2 7B	1074 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen2.5-Coder 7B	1089 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen2.5 14B	600 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen3 30B-A3B MoE	1142 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen2.5 32B	242 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Llama 3.1 70B	81.60 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen3-Coder-Next 80B-A3B	531 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	GPT-OSS 120B	120 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	MiniMax M2.1-REAP-139B	203 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	MiniMax M2.5	156 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen3-235B-A22B	101 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Nemotron-3 Nano 30B-A3B	112 tok/s	120 W	2026-03-03 (published)	visorcraft/strix-halo-llm-perf README (single-host table)

Adam Taylor (17)

Tester's site · Independent tester. Method: Python harness, 6 RTL tasks x 3 runs, temp 0.2, thinking off, 16K context, KV cache q8_0, models fully GPU-resident; generation rate as stated in post. Number as published in the post; no text copied.

Results from Adam Taylor
Platform	Workload	Model	Result	Power	Date	Original
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Qwen3.8 27B (dense)	18.90 tok/s	120 W	2026-09-23 (published)	Adam Taylor (Adiuvo Engineering) (manual extraction)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Qwen3-Coder 30B (MoE, 3.3B active)	47.50 tok/s	120 W	2026-09-23 (published)	Adam Taylor (Adiuvo Engineering) (manual extraction)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Qwen3.8-Flash-Next 125B (MoE, 6B active)	47.80 tok/s	120 W	2026-09-23 (published)	Adam Taylor (Adiuvo Engineering) (manual extraction)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	bluepearl-rtl (custom model config)	20.28 tok/s	120 W	2026-09-28 (published)	Adam Taylor MZ679 local-AI RTL run two (generation metadata)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	bluepearl-rtl (custom model config)	0.385 s	120 W	2026-09-28 (published)	Adam Taylor MZ679 local-AI RTL run two (generation metadata)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Gemma 4 12B (FPGA-tuned, "gemma4-fpga")	22.51 tok/s	120 W	2026-09-28 (published)	Adam Taylor MZ679 local-AI RTL run two (generation metadata)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Gemma 4 12B (FPGA-tuned, "gemma4-fpga")	0.470 s	120 W	2026-09-28 (published)	Adam Taylor MZ679 local-AI RTL run two (generation metadata)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen 3.5 9B (FPGA-tuned, "qwen-fpga")	33.98 tok/s	120 W	2026-09-28 (published)	Adam Taylor MZ679 local-AI RTL run two (generation metadata)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen 3.5 9B (FPGA-tuned, "qwen-fpga")	2.10 s	120 W	2026-09-28 (published)	Adam Taylor MZ679 local-AI RTL run two (generation metadata)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen3-Coder 30B	53.15 tok/s	120 W	2026-09-28 (published)	Adam Taylor MZ679 local-AI RTL run two (generation metadata)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen3-Coder 30B	0.215 s	120 W	2026-09-28 (published)	Adam Taylor MZ679 local-AI RTL run two (generation metadata)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen 3.5 4B	52.52 tok/s	120 W	2026-09-28 (published)	Adam Taylor MZ679 local-AI RTL run two (generation metadata)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen 3.5 4B	1.40 s	120 W	2026-09-28 (published)	Adam Taylor MZ679 local-AI RTL run two (generation metadata)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen3.8-Flash-Next	60.34 tok/s	120 W	2026-09-28 (published)	Adam Taylor MZ679 local-AI RTL run two (generation metadata)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen3.8-Flash-Next	0.900 s	120 W	2026-09-28 (published)	Adam Taylor MZ679 local-AI RTL run two (generation metadata)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen3.8 27B	34.17 tok/s	120 W	2026-09-28 (published)	Adam Taylor MZ679 local-AI RTL run two (generation metadata)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen3.8 27B	7.49 s	120 W	2026-09-28 (published)	Adam Taylor MZ679 local-AI RTL run two (generation metadata)

BDTI (12)

Tester's site · BDTI-measured whole-module power (4 MX3 chips + support circuitry) via current/power instrumentation on a pre-release MX3 M.2 module supplied by MemryX; average of last 3 s (30 readings at 10/s); host Intel Core i9-13900K, Ubuntu 22.04; 14 TFLOPS mode

Results from BDTI
Platform	Workload	Model	Result	Power	Date	Original
MemryX MX3 M.2 module	image classification	EfficientNet-Lite4	752 FPS	9.50 W	2025-05-15 (published)	BDTI Independent Report on MemryX MX3 M.2 AI Accelerator Module (May 2025)
MemryX MX3 M.2 module	image classification	EfficientNet-Lite4	30.00 FPS	2.80 W	2025-05-15 (published)	BDTI Independent Report on MemryX MX3 M.2 AI Accelerator Module (May 2025)
MemryX MX3 M.2 module	image classification	EfficientNet-Lite4	60.00 FPS	3.10 W	2025-05-15 (published)	BDTI Independent Report on MemryX MX3 M.2 AI Accelerator Module (May 2025)
MemryX MX3 M.2 module	image classification	EfficientNet-Lite4	90.00 FPS	3.30 W	2025-05-15 (published)	BDTI Independent Report on MemryX MX3 M.2 AI Accelerator Module (May 2025)
MemryX MX3 M.2 module	object detection	YOLOv8s	223 FPS	10.80 W	2025-05-15 (published)	BDTI Independent Report on MemryX MX3 M.2 AI Accelerator Module (May 2025)
MemryX MX3 M.2 module	object detection	YOLOv8s	30.00 FPS	3.70 W	2025-05-15 (published)	BDTI Independent Report on MemryX MX3 M.2 AI Accelerator Module (May 2025)
MemryX MX3 M.2 module	object detection	YOLOv8s	60.00 FPS	4.70 W	2025-05-15 (published)	BDTI Independent Report on MemryX MX3 M.2 AI Accelerator Module (May 2025)
MemryX MX3 M.2 module	object detection	YOLOv8s	90.00 FPS	5.90 W	2025-05-15 (published)	BDTI Independent Report on MemryX MX3 M.2 AI Accelerator Module (May 2025)
MemryX MX3 M.2 module	other	OSNet-x1	1054 FPS	6.70 W	2025-05-15 (published)	BDTI Independent Report on MemryX MX3 M.2 AI Accelerator Module (May 2025)
MemryX MX3 M.2 module	other	OSNet-x1	30.00 FPS	2.70 W	2025-05-15 (published)	BDTI Independent Report on MemryX MX3 M.2 AI Accelerator Module (May 2025)
MemryX MX3 M.2 module	other	OSNet-x1	60.00 FPS	2.80 W	2025-05-15 (published)	BDTI Independent Report on MemryX MX3 M.2 AI Accelerator Module (May 2025)
MemryX MX3 M.2 module	other	OSNet-x1	90.00 FPS	2.90 W	2025-05-15 (published)	BDTI Independent Report on MemryX MX3 M.2 AI Accelerator Module (May 2025)

arXiv 2608.00927 authors (10)

Tester's site ·

Results from arXiv 2608.00927 authors
Platform	Workload	Model	Result	Power	Date	Original
NVIDIA Jetson AGX Orin 64GB	object detection	YOLO11s (640x640)	80.90 FPS	60.00 W	2026-08-02 (published)	arXiv 2608.00927 AgriJetsonBench (abstract figures)
NVIDIA Jetson AGX Orin 64GB	object detection	YOLO11s (640x640)	0.181 J/inference	60.00 W	2026-08-02 (published)	arXiv 2608.00927 AgriJetsonBench (abstract figures)
NVIDIA Jetson AGX Orin 64GB	semantic segmentation	SegFormer-B0 (640x640)	31.95 FPS	60.00 W	2026-08-02 (published)	arXiv 2608.00927 AgriJetsonBench (abstract figures)
NVIDIA Jetson AGX Orin 64GB	semantic segmentation	SegFormer-B0 (640x640)	0.501 J/inference	60.00 W	2026-08-02 (published)	arXiv 2608.00927 AgriJetsonBench (abstract figures)
NVIDIA Jetson AGX Orin 64GB	semantic segmentation	SegFormer-B0 (640x640)	104 FPS	60.00 W	2026-08-02 (published)	arXiv 2608.00927 AgriJetsonBench (abstract figures)
NVIDIA Jetson Orin Nano Super	object detection	YOLO11s (640x640)	146 FPS	25.00 W	2026-08-02 (published)	arXiv 2608.00927 AgriJetsonBench (abstract figures)
NVIDIA Jetson Orin Nano Super	object detection	YOLO11s (640x640)	0.0767 J/inference	25.00 W	2026-08-02 (published)	arXiv 2608.00927 AgriJetsonBench (abstract figures)
NVIDIA Jetson Orin Nano Super	semantic segmentation	SegFormer-B0 (640x640)	47.18 FPS	25.00 W	2026-08-02 (published)	arXiv 2608.00927 AgriJetsonBench (abstract figures)
NVIDIA Jetson Orin Nano Super	semantic segmentation	SegFormer-B0 (640x640)	0.281 J/inference	25.00 W	2026-08-02 (published)	arXiv 2608.00927 AgriJetsonBench (abstract figures)
NVIDIA Jetson Orin Nano Super	semantic segmentation	SegFormer-B0 (640x640)	67.47 FPS	25.00 W	2026-08-02 (published)	arXiv 2608.00927 AgriJetsonBench (abstract figures)

From the Matrix (fromthematrix.dev) (9)

Tester's site · llama.cpp b8123 Vulkan, router mode; summarise a 7,624-token document; generation tok/s as published

Results from From the Matrix (fromthematrix.dev)
Platform	Workload	Model	Result	Power	Date	Original
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	LFM2-24B-A2B	109 tok/s	120 W	2026-03-05 (published)	From the Matrix (fromthematrix.dev) (manual extraction)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	gpt-oss-20b	70.00 tok/s	120 W	2026-03-05 (published)	From the Matrix (fromthematrix.dev) (manual extraction)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Nemotron-3-Nano-30B	61.00 tok/s	120 W	2026-03-05 (published)	From the Matrix (fromthematrix.dev) (manual extraction)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Qwen3-Coder-30B-A3B	66.00 tok/s	120 W	2026-03-05 (published)	From the Matrix (fromthematrix.dev) (manual extraction)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Qwen3.5-35B-A3B-UD	49.00 tok/s	120 W	2026-03-05 (published)	From the Matrix (fromthematrix.dev) (manual extraction)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Qwen3.5-35B-A3B	39.00 tok/s	120 W	2026-03-05 (published)	From the Matrix (fromthematrix.dev) (manual extraction)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	GLM-4.7-Flash	48.00 tok/s	120 W	2026-03-05 (published)	From the Matrix (fromthematrix.dev) (manual extraction)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	gpt-oss-120b	52.00 tok/s	120 W	2026-03-05 (published)	From the Matrix (fromthematrix.dev) (manual extraction)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Devstral-Small-2	14.00 tok/s	120 W	2026-03-05 (published)	From the Matrix (fromthematrix.dev) (manual extraction)

arXiv 2606.11257 authors (6)

Tester's site · Dell XPS 13 9345 (Snapdragon X Elite X1E80100, 64 GB), AC power; RAG queries, mean over 120 queries (wiki_minirag); NPU via QAIRT/QNN, CPU/GPU via GGUF Q4_K_M

Results from arXiv 2606.11257 authors
Platform	Workload	Model	Result	Power	Date	Original
Dell XPS 13 9345 (Snapdragon X Elite X1E80100)	other	Qwen3-4B-Instruct	14.19 tok/s	not reported	2026-06-09 (published)	arXiv 2606.11257 authors (manual extraction)
Dell XPS 13 9345 (Snapdragon X Elite X1E80100)	other	Qwen3-4B-Instruct	8.17 tok/s	not reported	2026-06-09 (published)	arXiv 2606.11257 authors (manual extraction)
Dell XPS 13 9345 (Snapdragon X Elite X1E80100)	other	Qwen3-4B-Instruct	4.65 tok/s	not reported	2026-06-09 (published)	arXiv 2606.11257 authors (manual extraction)
Dell XPS 13 9345 (Snapdragon X Elite X1E80100)	other	Qwen3-4B-Instruct	787 tok/s	not reported	2026-06-09 (published)	arXiv 2606.11257 authors (manual extraction)
Dell XPS 13 9345 (Snapdragon X Elite X1E80100)	other	Qwen3-4B-Instruct	43.36 tok/s	not reported	2026-06-09 (published)	arXiv 2606.11257 authors (manual extraction)
Dell XPS 13 9345 (Snapdragon X Elite X1E80100)	other	Qwen3-4B-Instruct	25.18 tok/s	not reported	2026-06-09 (published)	arXiv 2606.11257 authors (manual extraction)

Darren Soothill (Soot / Silicon) (6)

Tester's site · llama-bench pp512/tg128 on GMKtec EVO-X3 (Strix Halo); Vulkan build b10216 5 runs, ROCm build b10085 50 runs (revisions differ, disclosed)

Results from Darren Soothill (Soot / Silicon)
Platform	Workload	Model	Result	Power	Date	Original
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Qwen3-Coder-30B-A3B-Instruct	97.73 tok/s	120 W	2026-08-03 (published)	Darren Soothill (Soot / Silicon) (manual extraction)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Qwen3-Coder-30B-A3B-Instruct	73.65 tok/s	120 W	2026-08-03 (published)	Darren Soothill (Soot / Silicon) (manual extraction)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Qwen3.5-122B-A10B	21.30 tok/s	120 W	2026-08-03 (published)	Darren Soothill (Soot / Silicon) (manual extraction)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	LLM generation	Qwen3.5-122B-A10B	22.90 tok/s	120 W	2026-08-03 (published)	Darren Soothill (Soot / Silicon) (manual extraction)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen3-Coder-30B-A3B-Instruct	1115 tok/s	120 W	2026-08-03 (published)	Darren Soothill (Soot / Silicon) (manual extraction)
AMD Ryzen AI Max+ 395 / PRO 395 (Strix Halo, Radeon 8060S)	other	Qwen3-Coder-30B-A3B-Instruct	1345 tok/s	120 W	2026-08-03 (published)	Darren Soothill (Soot / Silicon) (manual extraction)

Mario Bergeron (5)

Tester's site · INA228 on the M.2 3.3 V rail (mb-powermon.py), average module power during benchmark runs; host AMD Ryzen AI Max+ 395

Results from Mario Bergeron
Platform	Workload	Model	Result	Power	Date	Original
Axelera Metis M.2 card	image classification	ResNet-50 v1.5	2050 FPS	7.60 W	2026-05-07 (published)	Mario Bergeron, Edge AI Power Benchmarking
DEEPX DX-M1 M.2 LPDDR5x2 module	image classification	ResNet-50	1009 FPS	4.70 W	2026-06-10 (published)	Mario Bergeron, Edge AI Power Benchmarking
Hailo-8 M.2 module	image classification	ResNet-50	1371 FPS	4.00 W	2026-05-05 (published)	Mario Bergeron, Edge AI Power Benchmarking
MemryX MX3 M.2 module	image classification	ResNet-50	1796 FPS	11.00 W	2026-06-19 (published)	Mario Bergeron, Edge AI Power Benchmarking
MemryX MX3 M.2 module	image classification	ResNet-50	2335 FPS	20.20 W	2026-06-19 (published)	Mario Bergeron, Edge AI Power Benchmarking

arXiv 2603.23640 authors (4)

Tester's site · Qwen2.5 1.5B 4-bit, fixed 258-token prompt, greedy decoding, 20 warm iterations; mean of runs 2-20; system power measured (100 ms polling)

Results from arXiv 2603.23640 authors
Platform	Workload	Model	Result	Power	Date	Original
Apple iPhone 16 Pro (A18 Pro)	LLM generation	Qwen2.5 1.5B (Q4)	25.89 tok/s	not reported	2026-06-07 (published)	arXiv 2603.23640 authors (manual extraction)
Laptop with NVIDIA RTX 4050 Laptop GPU (Core i5-13420H, on battery)	LLM generation	Qwen2.5 1.5B (Q4)	132 tok/s	34.12 W	2026-06-07 (published)	arXiv 2603.23640 authors (manual extraction)
Raspberry Pi 5 + Hailo-10H (AI HAT+ 2 class, PCIe Gen3)	LLM generation	Qwen2.5 1.5B (Q4)	6.91 tok/s	1.87 W	2026-06-07 (published)	arXiv 2603.23640 authors (manual extraction)
Samsung Galaxy S24 Ultra (Snapdragon 8 Gen 3)	LLM generation	Qwen2.5 1.5B (Q4)	10.83 tok/s	not reported	2026-06-07 (published)	arXiv 2603.23640 authors (manual extraction)

Grades: A audited MLPerf · B third-party · C community · D vendor. How we rank · JSON