H200 Training Runs Smoothly | Generated by AI
Training is running smoothly on the H200. Here’s the status:
Model: d24 — 24 layers, 1536 dim, 12 heads, 1.38B params (config shows it’s actually bigger than the original 760M reference) FP8: enabled (145/158 layers converted)
Performance:
- ~129K tokens/sec — nearly 2× the MI300X (68K tok/s)
- ~62% MFU — much higher than MI300X (27%) thanks to FA3 + FP8 on H200
- ~4.1s per step
- ETA: ~33 hours (vs 62h on MI300X)
Training curve (looking good):
step 0: loss 10.40
step 10: loss 9.16
step 20: loss 7.21
Loss is dropping cleanly — healthy start. At this rate the full 29K steps should finish in ~33 hours, significantly faster than the original MI300X estimate.