Blackwell GPU Configuration for ML Nodes

[votkon]

Please note that after recent PoC updates the weight of B200 GPUs decreased considerably and is now barely equal to H100 or H200 weight performance.

PoC v2 inference-based weight adjustments

PoC v2 inference-based weight adjustments With PoC v2 active, weight assignment is now based on measured inference performance on the current model ⁨⁨⁨⁨⁨⁨⁨Qwen/Qwen3-235B-A22B-Instruct-2507-FP8⁩⁩⁩⁩⁩⁩⁩. As a result, both median GPU weights and relative weight ratios between GPU types have been adjusted. Observed GPU weight changes (Epoch 158 → 159) ┌────────────────┬───────────┬───────────┬────────┐ │ GPU Type │ Epoch 158 │ Epoch 159 │ Change │ ├────────────────┼───────────┼──────…