Default load on 2×+ DGX Spark desks. EXL3 recipes on GitHub. No measured decode yet.
Params
unknown
Fastest decode
264 tok/s
4× RTX PRO 6000 · EXL3 · TensorFold 0.6.2
Refusal
Official
Setups
13
Best reported pair · not live serving speed ↗
Runs on
Quoted one-stream decode on real boxes.
- 4× RTX PRO 6000EXL3264 tok/s
- 2× RTX PRO 6000EXL3238 tok/s
- 4× DGX Sparkunknown127 tok/s
- RTX PRO 6000 + 2× GB10EXL3 4bpw120 tok/s
- 4× DGX SparkEXL3 4bpw104 tok/s
- 2x DGX SparkEXL385 tok/s
- 2× DGX SparkEXL360.4 tok/s
- 2x DGX SparkNVFP455.6 tok/s
- DGX SparkEXL3 2.05bpw32.3 tok/s
- DGX SparkEXL3 2.05bpw30 tok/s
- DGX SparkEXL318.6 tok/s
- 2x DGX SparkNVFP415 tok/s
- Mac Studio M4 Max 64GBQ211.3 tok/s