Performance tuning the NVIDIA Grace CPU, part of both the NVIDIA Grace CPU Superchip and NVIDIA Grace Hopper Superchip, ensures that your applications make the most of the NVIDIA data center computing platform. Key takeaways include:
Hardware performance analysis: Arm Performance Monitoring Units (PMUs) on the Grace CPU enable performance analysis tools to capture and characterize hardware performance. Arm Statistical Profiling Extensions (SPE) provide low-overhead, high-resolution, hardware-based statistical sampling for performance measurement.
Nsight Systems 2023.4 Grace CPU event sampling: Grace CPU core event rates show how instructions are executed on Grace CPU cores. Available now, Nsight Systems Grace CPU uncore event sampling reveals CPU activity outside of the cores, including PCIe events and NvLink-C2C activity.
CUDA kernel performance: An important part of NVIDIA data center performance tuning is optimizing CUDA code. Nsight Compute is the NVIDIA tool for inspecting CUDA performance. Performance monitor sampling (PM sampling) reveals time-correlated data for CUDA kernel performance.
Source code comparison: Nsight Compute 2023.3 source code comparison enables you to inspect how source-level modifications impact CUDA kernel performance. Automatically compare source files and correlate changes with associated performance metrics.
00:00 - Introduction
0:26 - Grace CPU Performance Monitoring
0:47 - Nsight Systems Grace CPU Sampling
1:02 - Nsight Systems Core Event Rates
1:20 - Nsight Systems Uncore Event Rates
2:41 - Nsight Compute PM Sampling
3:00 - Nsight Compute Source Code Comparison
Read the NVIDIA Grace CPU Superchip whitepaper: [ Ссылка ]
Check out tutorials for using NVIDIA Nsight Systems and NVIDIA Nsight Compute: [ Ссылка ]
Get started with NVIDIA Nsight Systems: [ Ссылка ]
Get started with NVIDIA Nsight Compute: [ Ссылка ]
Learn more about Nsight Developer Tools: [ Ссылка ]
Join the NVIDIA Developer Program: [ Ссылка ]
Read and subscribe to the NVIDIA Technical Blog: [ Ссылка ]
#NVIDIAGraceCPU #CUDA #Nsight #developertools #NVIDIA #HPC
Ещё видео!