Measurements pending. CPU-only and Metal inference have been tried, but no reproducible numeric dataset is available for publication. This page does not report live performance.
| Metric | CPU-only | Metal / Vega 64 | Unit |
|---|---|---|---|
| Prompt processing | Not measured for publication | Not measured for publication | tokens / second |
| Output generation | Not measured for publication | Not measured for publication | tokens / second |
| Time to first token | Not measured for publication | Not measured for publication | seconds |
| Peak memory | Not measured for publication | Not measured for publication | GB |
| Task quality | Not evaluated for publication | Not evaluated for publication | task-specific rubric |
Proposed measurement method
Record the exact model, quantization, software build, context length, offload settings, and sampling configuration. Run fixed workloads repeatedly and separate startup effects from sustained operation.
Report prompt processing and token generation independently. Include output quality and failed runs so speed does not become a substitute for usefulness.
Current observation
The owner reports that GPU acceleration on the Vega 64 has not reliably improved performance. Without matched configurations and repeated measurements, that observation should not become a general hardware ranking.
Read the measurement plan