MI300X 760M Model Training Estimate | Generated by AI
Estimated training time for 760M on 14.6B tokens, single MI300X: ~24-35 hours
Here’s the breakdown:
Total FLOPs: 6.66 × 10^19
MI300X peak: 1,307 TFLOPS (bf16)
Realistic MFU: 40-60% (single GPU, no communication overhead)
At 40% MFU: ~35 hours
At 50% MFU: ~28 hours ← most likely
At 60% MFU: ~24 hours
Cross-validation with Karpathy’s nanochat:
- d24 (~1.5B) on 8×H100: ~2 hours for 4.64e18 FLOPs
- Our run: 6.66e19 FLOPs = 14× more compute
- On 8×H100: 14 × 2 = 28 hours
- On 1×MI300X (1.33× faster than H100): ~21 hours
- Account for single-GPU overhead: ~24-28 hours
Comparison with your 124M model on RTX 4070:
- 124M on 4070, 10B tokens: ~15-20 hours
- 760M is 6.1× more FLOPs/token
- MI300X is ~10.5× faster (bandwidth-bound case)
- Net: 15-20 × 6.1 × (14.6/10) / 10.5 ≈ 13-17 hours (if bandwidth-limited)
- But 760M is compute-bound → closer to 24-28 hours
Bottom line: ~1-1.5 days on a single MI300X. Not bad for 14.6B tokens through a 760M model. You could start it in tmux and come back tomorrow.
Generated by AI. Curating and sharing still takes effort.
If you find it useful, feel free to
donate.
WeChat: @lzwjava
·
X: @lzwjava ·
Say hi 👋