Framework Desktop Mainboard and 4-node cluster (as tested by Jeff Geerling): edge AI benchmarks
- Independent test Jeff Geerling
Key results
How to read power: results are compared only within the same model. By efficiency ranks only results whose source publishes a power figure; By speed ranks every result. Measured = measured for this workload. Vendor-measured (typical/avg) = the vendor’s own measurement at typical or average load, not an independent test and not a maximum, so fps/W looks better than at max TDP. Chip, Module or Board TDP = the vendor’s published maximum, not measured. Typical power, not max TDP = only a typical figure is published, so fps/W looks better than at full load. Scope says what the watts cover: Accelerator only (host not included), Module, or Whole system. Max of range = top of a published range, so fps/W is a floor (≥). “— not published” = the source gives no power; we never estimate one.
| Workload | Model · precision | Result | Power | Perf/W | Rank on this model | Grade · date | Source |
|---|
Ranks compare this board only with boards running the same model. Up to three models per workload; open a rank to see the full same-model table.
Show data — per-workload tables
object detection
No published object detection results for Framework Desktop Mainboard and 4-node cluster (as tested by Jeff Geerling) yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.
instance segmentation
No published instance segmentation results for Framework Desktop Mainboard and 4-node cluster (as tested by Jeff Geerling) yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.
pose estimation
No published pose estimation results for Framework Desktop Mainboard and 4-node cluster (as tested by Jeff Geerling) yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.
image classification
No published image classification results for Framework Desktop Mainboard and 4-node cluster (as tested by Jeff Geerling) yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.
semantic segmentation
No published semantic segmentation results for Framework Desktop Mainboard and 4-node cluster (as tested by Jeff Geerling) yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.
speech-to-text
No published speech-to-text results for Framework Desktop Mainboard and 4-node cluster (as tested by Jeff Geerling) yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.
LLM generation
| Like-for-like group | Best result | Model · precision · stack | Grade | Power | Source · date | Rank in group |
|---|---|---|---|---|---|---|
| LLM ≤4B · b=1 | 85.38 tok/s CPU-only | deepseek-r1:1.5b Ollama default tag (4-bit unless tag says otherwise) · Ollama (geerlingguy/ollama-benchmark) | C | 142 W ● Measured, Whole system | Jeff Geerling ai-benchmarks issue results (comments) 2025-08-07 (page last edited) | 38 of 81 |
| LLM 5–14B · b=1 | 20.89 tok/s CPU-only | llama3.1:8b Ollama default tag (4-bit unless tag says otherwise) · Ollama (geerlingguy/ollama-benchmark) | C | 142 W ● Measured, Whole system | Jeff Geerling ai-benchmarks issue results (comments) 2025-08-07 (page last edited) | 45 of 66 |
| LLM >40B · b=1 | 2.43 tok/s CPU-only | deepseek-r1:70b Ollama default tag (4-bit unless tag says otherwise) · Ollama (geerlingguy/ollama-benchmark) | C | 141 W ● Measured, Whole system | Jeff Geerling ai-benchmarks issue results (comments) 2025-08-07 (page last edited) | 10 of 13 |
Grades: A audited MLPerf · B third-party · C community · D vendor. How we rank · JSON