INFERENTIA 1 vs TRAINIUM 2
The TRAINIUM 2 has 96 GB of memory against 8 GB on the INFERENTIA 1.
| Measure | INFERENTIA 1 | AWS Trainium2 |
|---|---|---|
| Lowest per GPU-hour | — | $7.49 |
| Median per GPU-hour | — | $7.49 |
| On-demand from | — | — |
| Spot from | — | $7.49 |
| Reserved from | — | — |
| Lowest per A100-equivalent hour | — | $3.38 |
| Live offers | 0 | 1 |
| Providers | 0 | 1 |
| Memory | 8 GB DDR4 | 96 GB HBM3 |
| Memory bandwidth | — | 2,900 GB/s |
| FP16 / BF16 | 64 TFLOPS | 667 TFLOPS |
| FP8 | — | 1,299 TFLOPS |
| Performance vs A100 80GB | 0.14× | 2.22× |
| Architecture | Inferentia (NeuronCore v1) | Trainium2 (NeuronCore v3) |
| Power | — | 500 W |
Where the offers sit
No live offers right now.