HiSilicon AI Accelerators

Ascend AI accelerators feature: 1) Dedicated AI cores optimized for neural network inference and training

2) Industry-leading TOPS-per-watt performance for power-efficient AI processing

3) Support for popular AI frameworks including TensorFlow, PyTorch, and MindSpore

4) Scalable architecture from edge devices to data center servers

5) Comprehensive software stack with compilers, optimizers, and profiling tools

6) Hardware security features for AI model protection. These features make Ascend ideal for applications requiring real-time AI processing with strict power constraints.

Review the specifications to identify the Ascend model that meets your AI performance and power requirements.

Ascend 310 and 910 serve different application domains: Ascend 310 is designed for edge AI applications with 16 TOPS INT8 performance and 8W typical power consumption, ideal for cameras, robots, and edge gateways. Ascend 910 is a data center AI processor delivering 256-512 TFLOPS FP16 performance with 310W power consumption, designed for training large AI models. Choose Ascend 310 for inference at the edge where power and cost are critical, and Ascend 910 for data center training and high-throughput inference. Both share the same Da Vinci architecture and software stack for easy migration.

Define your deployment location (edge vs. data center) and AI workload type (inference vs. training) to select the appropriate Ascend model.

Ascend AI accelerators support a comprehensive range of AI frameworks: 1) MindSpore - Huawei's native AI framework with optimal Ascend optimization

2) TensorFlow and TensorFlow Lite - via graph conversion tools

3) PyTorch - through ONNX export and conversion

4) ONNX - direct support for standardized models

5) Caffe and Caffe2 - via model conversion

6) Custom operators - through CANN (Compute Architecture for Neural Networks) SDK. The CANN software stack provides model conversion, optimization, and deployment tools to run models from any framework efficiently on Ascend hardware.

Use MindSpore for best performance, or convert your existing TensorFlow/PyTorch models using the CANN conversion tools.

Ascend AI accelerator power consumption varies by model: Ascend 310 (edge) consumes 8W typical with 16 TOPS INT8 performance, achieving 2 TOPS/W efficiency. Ascend 310P (enhanced edge) consumes 15W with 22 TOPS performance. Ascend 910 (data center) consumes 310W with 256 TFLOPS FP16 performance, delivering excellent training efficiency. Ascend 910B (enhanced) reaches 376 TFLOPS at 310W. All models include dynamic power management and can operate in reduced power modes when full performance is not required.

Select Ascend 310 series for edge applications with strict power budgets, and Ascend 910 series for data center applications.

HiSilicon provides comprehensive development tools for Ascend: 1) CANN (Compute Architecture for Neural Networks) - core software stack including drivers, runtime, and libraries

2) MindStudio - integrated development environment for model development and optimization

3) Model conversion tools - convert TensorFlow, PyTorch, and ONNX models to Ascend format

4) Profiling tools - analyze model performance and identify bottlenecks

5) SDK and APIs - for custom operator development

6) Reference designs and sample applications. BeiLuo provides technical support and training for Ascend development.

Contact BeiLuo to access Ascend development tools and receive training on the CANN software stack.