Side by side datacenter GPU specifications from the Flopper.io database.
| Huawei Atlas 350 | Huawei Ascend 950PR | |
|---|---|---|
| FP32 TFLOPS | — | — |
| Memory | 112 GB | 128 GB |
| Memory Bandwidth | 1.4 TB/s | 1.6 TB/s |
| TDP | 600 W | 900 W |
| Precision | Huawei Atlas 350 peak TFLOPS (dense) | Huawei Ascend 950PR peak TFLOPS (dense) |
|---|---|---|
| FP64 | — | — |
| FP32 | — | — |
| TF32 | — | — |
| BF16 | 425.0 | 500.0 |
| FP16 | 425.0 | 500.0 |
| FP8 | 804.0 | 1000.0 |
| FP6 | 804.0 | — |
| FP4 | 1561.0 | 2000.0 |
| INT8 | 804.0 | — |
| Huawei Atlas 350 | Huawei Ascend 950PR | |
|---|---|---|
| Architecture | Da Vinci v3 | Da Vinci v3 |
| Process Node | — | — |
| Launch Year | 2026 | 2026 |
| Form Factor | Custom | OAM |
| Memory Type | HiBL 1.0 | HiBL 1.0 |
| Interface Width | — | — |
| Interconnect | Unified Bus 2.0 | Unified Bus 2.0 |
| TDP | 600 W | 900 W |
| FP32 Efficiency | — | — |