Skip to main content

Huawei Ascend

Vendor: Huawei

Category: NPU Neural Processor

Architecture: Da Vinci

Introduction​

Huawei Ascend AI processor series, including training chip Ascend 910B/910 and inference chip Ascend 310P/310. Paired with CANN (Huawei AI Compute Framework) and MindSpore deep learning framework, widely used in domestic AI infrastructure construction.

Specifications​

ModelComputeMemoryInterfaceTDPProcess
Ascend 910B280 TFLOPS (FP16) / 560 TOPS (INT8)64GB HBM2eOAM310W7nm
Ascend 910256 TFLOPS (FP16)32GB HBM2PCIe 4.0300W7nm
Ascend 310P70 TOPS (INT8)24GB LPDDR4XPCIe 4.075W12nm
Ascend 31022 TOPS (INT8)8GB LPDDR4PCIe 3.08W12nm

Official Website​

Visit Official Website

Driver Downloads​

Linux​

OS Support​

WindowsLinuxmacOSAndroid
❌✅❌❌

Version History​

VersionRelease DateDescription
CANN 8.02024-Q4910B comprehensive optimization + MindSpore 2.x
CANN 7.02024-Q1torch_npu PyTorch native adaptation
CANN 6.02023-Q13D Cube architecture deep optimization

Performance Benchmarks​

ModelTaskPerformance Metric
Ascend 910B × 8GPT-3 175B Training~1.5 days (official data)
Ascend 910BLlama 2 70B Inference~80 tok/s (FP16)
Ascend 310PImage Classification/Detection~2000 img/s (ResNet-50)
Ascend 310Edge Inference8W ultra-low power

Pricing​

ModelReference PriceNotes
Ascend 910BContact vendorPrimarily sold through Atlas product line
Ascend 310PContact vendorAtlas 300I Pro
Ascend 310Contact vendorAtlas 200 DK developer kit

Quick Installation​

Linux (Ubuntu 22.04 / EulerOS)​

# 1. Install Ascend driver and firmware
sudo ./Ascend-cann-npu-driver_*.run --install

# 2. Install CANN toolkit
sudo ./Ascend-cann-toolkit_*.run --install

# 3. Set environment variables
source /usr/local/Ascend/ascend-toolkit/set_env.sh

# 4. Verify installation
npu-smi info

Code Examples​

Python (torch_npu)​

import torch
import torch_npu  # Ascend NPU extension

# Use NPU backend
device = torch.device("npu:0")
x = torch.randn(1024, 1024, device=device)
y = torch.matmul(x, x)
print(f"NPU matrix multiply: {y.shape}")
print(f"NPU device: {torch.npu.get_device_name(0)}")

MindSpore (Ascend Native Framework)​

import mindspore as ms

# MindSpore automatically uses Ascend NPU
context.set_context(device_target="Ascend")
x = ms.Tensor(np.random.randn(1024, 1024), ms.float32)
y = ms.ops.matmul(x, x)
print(f"MindSpore NPU multiply: {y.shape}")

Architecture Highlights​

  • Da Vinci Architecture: 3D Cube compute unit, optimized for matrix multiplication; Cube units accelerate Transformer attention computation
  • CANN Full Stack: Compute Architecture for Neural Networks — complete software stack from driver to compiler to operator library
  • torch_npu: PyTorch's Ascend NPU backend, API highly consistent with CUDA backend, low migration cost
  • MindSpore: Huawei proprietary full-scenario AI framework, automatic differentiation + distributed training

Model Compatibility​

Model/FrameworkSupportNotes
MindSpore✅ NativeBest support
PyTorch✅ torch_npuCUDA-API compatible
TensorFlow⚠️Via CANN adaptation
Llama / Qwen and similar LLMs✅MindIE / torch_npu both supported
PaddlePaddle⚠️Under adaptation
Speech/Vision Models✅ModelScope support

Large-Scale Cluster Deployments​

Based on global AI supercomputing cluster statistics, Huawei Ascend has accumulated over 6,000 chips deployed across 1 publicly disclosed cluster.

Chip Model Statistics​

Chip ModelTotal DeployedCluster Count
Huawei Ascend 910B6,0001

Notable Deployment Clusters Top 10​

#Cluster NameTotal ChipsChip ModelOperator
1Huawei Pangu Ultra MoE 910Bs6,000Huawei Ascend 910B ×6,000

If you're evaluating alternatives, the following products may also fit your scenario: