Research on DeepSeek-Oriented Benchmarking Methods for AI Inference Accelerators

With the rapid shift of large model technology focus from training to inference, evaluating the performance of domestic AI accelerators in inference scenarios has become critical for building intelligent computing infrastructure. To tackle this, this paper proposes a multi-dimensional evaluation methodology for benchmarking AI accelerator inference performance, based on the DeepSeek-R1 series of models. Through quantitative analysis of how hardware parameters (computing power, memory bandwidth, interconnect bandwidth) and quantization schemes impact inference efficiency, this study provides empirical evidence to support the selection of domestic hardware options. Experimental results indicate that memory bandwidth is a primary bottleneck for large model decoding throughput, while INT8 quantization enables approximately 2x throughput gains compared to FP16 precision. This research offers valuable insights for hardware selection and performance optimization in intelligent computing centers, thereby enhancing the competitiveness of domestic AI accelerators in the inference market.