ASCEND 910B3 vs ASCEND 910B4
The cheapest ASCEND 910B4 ($0.31) costs 40% less per hour than the cheapest ASCEND 910B3 ($0.51).
Per unit of performance, the ASCEND 910B4 is 15% cheaper: $0.48 versus $0.56 per A100-equivalent hour.
The ASCEND 910B3 has 64 GB of memory against 32 GB on the ASCEND 910B4.
| Measure | ASCEND 910B3 | ASCEND 910B4 |
|---|---|---|
| Lowest per GPU-hour | $0.51 | $0.31 |
| Median per GPU-hour | $0.51 | $0.40 |
| On-demand from | — | — |
| Spot from | — | — |
| Reserved from | $0.51 | $0.31 |
| Lowest per A100-equivalent hour | $0.56 | $0.48 |
| Live offers | 1 | 2 |
| Providers | 1 | 2 |
| Memory | 64 GB HBM2e | 32 GB HBM2e |
| Memory bandwidth | 1,600 GB/s | 800 GB/s |
| FP16 / BF16 | 313 TFLOPS | 280 TFLOPS |
| FP8 | — | — |
| Performance vs A100 80GB | 0.91× | 0.65× |
| Architecture | 达芬奇 (Da Vinci) / 昇腾 910B | 达芬奇 (Da Vinci) / 昇腾 910B |
| Power | 350 W | 300 W |