GPUFlight - GPU Profiler & Monitoring System Blog https://gpu-flight.github.io/blog GPUFlight - GPU Profiler & Monitoring System Blog Tue, 01 Apr 2025 00:00:00 GMT https://validator.w3.org/feed/docs/rss2.html https://github.com/jpmonette/feed en <![CDATA[AMD ROCm Support]]> https://gpu-flight.github.io/blog/amd-rocm-support https://gpu-flight.github.io/blog/amd-rocm-support Tue, 01 Apr 2025 00:00:00 GMT GPUFlight now supports AMD GPUs via ROCm, including HIP kernel tracing, system telemetry, occupancy analysis, ISA disassembly, and extended hardware metrics.

What's New​

  • AMD/ROCm backend: Kernel tracing via rocprofiler-sdk, system metrics via ROCm SMI
  • Occupancy analysis: Theoretical occupancy computed from VGPR usage, LDS, and wavefront limits
  • ISA disassembly: Automatic RDNA ISA capture via llvm-objdump
  • Extended metrics: Fan speed, junction/memory temperature, voltage, energy, PCIe bandwidth, ECC errors
  • Report generation: gpufl::generateReport() produces a text summary after profiling
  • CPU iGPU filtering: AMD APU integrated graphics are automatically excluded from telemetry
  • Vendor-neutral frontend: Web UI adapts terminology (SASS/ISA, Warp/Wavefront) based on GPU vendor

Supported Hardware​

  • NVIDIA: CUDA Toolkit with CUPTI (unchanged)
  • AMD: ROCm 6.x with HIP, rocprofiler-sdk, and ROCm SMI

See the AMD Integration Guide for details.

]]> Release AMD Feature