Real end-to-end training runs — loss actually decreases, not smoke tests. What trains on stock Trainium and what needs Luma:
| # | Family | Model | Type | Stock Trainium | With Luma |
|---|---|---|---|---|---|
| 1 | Decoder LLM | distilgpt2 | Text-to-text | ✕ NaN | ✓ trains |
| 2 | Encoder | distilbert-base-uncased | Text encoder (fill-mask) | ✓ trains | ✓ trains |
| 3 | Encoder-decoder | t5-small | Text-to-text | ✓ trains | ✓ trains |
| 4 | ViT | vit-base-patch16-224 | Image classification | ✓ trains | ✓ trains |
| 5 | CNN | resnet-18 | Image classification | ✓ trains | ✓ trains |
| 6 | Diffusion UNet | ddpm-cifar10-32 | Image generation | ✕ no converge | ✓ trains |
| 7 | STT | whisper-tiny | Speech-to-text | ✓ trains | ✓ trains |
| 8 | VLA | smolvla_base | Vision-language-action | ✕ compile err | ✓ trains |
| 9 | VLM | SmolVLM-256M-Instruct | Image-text-to-text | ✕ crash | ✓ trains |
| 10 | MoE | switch-base-8 | Text-to-text (MoE) | ✕ hangs | ✓ trains |
Method: each model is trained end-to-end on a trn1 (forward + backward + optimizer); ✓ means the loss actually decreases, not merely that the graph compiles. Stock Trainium is unmodified Neuron — encoder, seq2seq, vision and speech models train as-is, while causal LLM (NaN), MoE (hangs), diffusion (no convergence), VLM (crash) and VLA do not. With Luma closes every gap — all ten train end-to-end (smolvla's stock-Neuron result is still pending a clean bare run).
🟩
You don't need us here.
Everything already runs.
Where AMD Instinct wins:
| # | Family | Model | Type | Parameters | Runs on AMD Instinct |
|---|---|---|---|---|---|
| 1 | LLM | ~120B dense | Inference (fp16) | 120B (~240 GB) | ✅ one MI325X — Nvidia needs 2× B200/H200 + TP |
| 2 | Long-context | large KV cache | Inference | > 192 GB | ✅ fits — Nvidia capped at 192 GB |
| 3 | Big-batch | 70B–200B | Inference (memory-bound) | 70B–200B | ✅ MI300X ~$15–25K — B200's 192 GB at ~⅓ price |