HiSilicon AI Accelerator Selection Guide
This technical reference document provides detailed information about hisilicon product specifications, characteristics, and performance parameters. Use this information to support your design and analysis activities.
Electrical characteristics are specified over the operating temperature range unless otherwise noted. Parameters are guaranteed by design, testing, or statistical analysis. Typical values is the most likely parametric norm at 25°C.
Thermal characteristics require careful attention during system design. The junction-to-ambient thermal resistance depends on the mounting configuration, PCB copper area, and airflow conditions. Use thermal simulation tools to predict operating temperatures under actual conditions.
Reliability data is based on accelerated life testing and field failure analysis. Mean time between failures (MTBF) calculations follow industry-standard methodologies. Contact BeiLuo for detailed reliability reports and qualification data.
💡 FAE Insights
📋 Customer Cases
Video Analytics Provider
Security
Challenge
Needed to process 10,000 camera feeds with real-time object detection
Solution
Deployed Ascend 310 edge nodes with Ascend 910 cluster for model training
Customer Feedback
"Customer reported successful implementation and improved performance."
Results
Achieved 50ms end-to-end latency with 95% detection accuracy
Frequently Asked Questions
1. How do I migrate my TensorFlow models to Ascend?
Migration involves: 1) Export your TensorFlow model to ONNX or SavedModel format; 2) Use CANN's ATC (Ascend Tensor Compiler) to convert to Ascend format; 3) Optimize using profiling tools; 4) Deploy using CANN runtime. Simple models convert in hours; complex models with custom ops may take 1-2 weeks. CANN provides automatic operator mapping for standard TensorFlow ops.
2. What is the TOPS-per-watt of Ascend processors?
Ascend 310 achieves 2 TOPS/W (16 TOPS at 8W), among the best in the industry for edge AI. Ascend 910 achieves approximately 0.83 TFLOPS/W for FP16 (256 TFLOPS at 310W), competitive with data center GPUs. The power efficiency makes Ascend attractive for both edge battery-powered devices and large data center deployments where power costs are significant.
3. Can Ascend 910 be used for inference as well as training?
Yes, Ascend 910 excels at both: Training - 256 TFLOPS FP16 for model training; Inference - 512 TOPS INT8 for high-throughput inference. Many customers use Ascend 910 for both, simplifying their AI infrastructure. For pure inference at lower cost, consider Ascend 310 for edge or multiple Ascend 310 cards for data center inference.
4. FAQ 4 for HiSilicon AI Accelerator Selection Guide
Please refer to the article content or contact our FAE team for detailed information. For detailed specifications and application support on hisilicon products, refer to the datasheet or contact our team.
5. FAQ 5 for HiSilicon AI Accelerator Selection Guide
Please refer to the article content or contact our FAE team for detailed information. For detailed specifications and application support on hisilicon products, refer to the datasheet or contact our team.