Measure GPU matrix computation throughput with standardized TFLOPs or TOPs metrics.
npx skills add https://github.com/dongg622/china-ai-chip-skill --skill gemm-perf
Please help me install this Agent Skill. Skill: gemm-perf Source: https://github.com/dongg622/china-ai-chip-skill/tree/main/TIANSHU/gemm-perf Command: npx skills add https://github.com/dongg622/china-ai-chip-skill --skill gemm-perf
This Skill enables hardware developers and FAEs to measure the matrix computation performance of GPU accelerators efficiently.
Run the gemm_perf tool directly on your system with the default parameters to measure your device's matrix computation speed.
High-intent search queries and answers about installing and using this skill.
To benchmark GPU matrix performance, you can run tests that evaluate GEMM computational throughput across data types like FP32, FP16, BF16, and INT8. This process outputs standardized TFLOPs or TOPs metrics for direct hardware comparison.
Testing GEMM capacity involves conducting comprehensive matrix performance tests on the GPU hardware. This validates whether the AI accelerator meets required specifications by measuring computational throughput and providing precise TFLOPs or TOPs metrics.
Yes, measuring GPU matrix computation speed requires compatible GPU drivers and the corex-toolbox. These prerequisites ensure the environment can accurately execute the performance tests and report standardized throughput metrics.
Yes, you can compare hardware validation results using standardized TFLOPs or TOPs metrics. The performance benchmarking tests output these standardized metrics to enable direct comparison of matrix computation speed across different GPU hardware.
Matrix performance testing handles multiple data formats by executing GEMM operations across supported types like FP32, FP16, BF16, and INT8. This allows comprehensive evaluation of computational throughput for each specific data format.
sandraschiSolve advanced linear algebra problems with LaTeX-formatted explanations and worked examples.
tunnox-netIdentify tunneling system bottlenecks and prescribe latency, throughput, and concurrency optimizations.
graphsignalProfile production inference workloads via sidecar process without modifying source code.
pacphiDefine and train PyTorch neural networks with CUDA acceleration.
