

Ascend
#10 en Cloud IA & MLOpsHuawei · depuis Ascend 910 (erste Generation): 23. August 2019; Ascend 310: November 2018 (laut Drittquellen) · 2× · vu le 17 sept. 2026
46
Momentum
Ascend is a family of AI accelerators (NPUs) and servers from Huawei for training and inference of large language models and other AI workloads. The chips are based on the Da Vinci architecture and run on the CANN software stack and the MindSpore framework. According to third-party sources, Huawei does not publish official list prices for Ascend chips or Atlas systems.
Historique du momentum
18.08.16.11.
Fonctionnalités
| Manufacturing Process (nm) | Ascend 910: 7 nm (N7+); Ascend 310: 12 nm; Ascend 950 series: 6 nm (press reports) |
| Platform | Data-center training (Ascend 910), edge inference (Ascend 310); Atlas servers, cloud instances via Huawei Cloud |
| Price | Not officially published by Huawei |
| Compute Performance (FLOPS/TOPS) | Ascend 910: 256 TFLOPS FP16 / 512 TOPS INT8; Ascend 310: 16 TOPS INT8 / 8 TFLOPS FP16; Ascend 910B: 280 TFLOPS FP16 / 560 TOPS INT8 |
| Release Date | Ascend 910: 23 August 2019 (Shenzhen); Ascend 310: November 2018 |
| Memory | Ascend 910: 32 GB HBM2; Ascend 910B: 64 GB HBM2e; Ascend 310P: 24 GB LPDDR4X |
| Availability | Ascend 950 only via Huawei Cloud China, Alibaba Cloud China and ByteDance Volcengine; Ascend 960DT announced for Q1 2027 |