HIP kernels, fully symbolized
Full kernel names, such as rocBLAS GEMM TN [bf16xbf16→bf16, accum=f32].
AMD GPU profiling
Always-on ROCm profiling
A slow kernel tells you little without the code that launched it. Zymtrace reveals the calling stack and instruction-level stalls, including for vLLM and SGLang with HIP graphs, so your team knows where to optimize.
- Fleet-wide
- No code changes
AMD SMI metrics
See the code behind any AMD GPU metric
Click any point on any AMD GPU chart to open the ROCm flamegraph of that moment.
- Utilization and memory activity per device
- Power against the power limit, temperature and clocks
- Read from AMD SMI on each host
ROCm coverage
One profiler for NVIDIA and AMD
Built on the ROCprofiler SDK, with the same CPU context as CUDA profiles.
Instructions and stall reasons
Instruction-level samples show where inside a kernel the time goes.
Memory and sync
Copies and synchronization as their own frames, next to compute.
ROCTX ranges
ROCTX ranges in the flamegraph, on by default.
Inference engines on ROCm
vLLM and SGLang, with HIP graphs and engine metrics.
One fleet view
AMD and NVIDIA hosts side by side in Efficiency IQ and Top Functions.
FAQ
AMD GPU profiling FAQ
AMD GPUs on x86_64 hosts with the amdgpu kernel driver and ROCm 7 or later, including the MI300X and MI325X.
Yes, once. ROCm profiling starts with the process, by setting
ROCP_TOOL_LIBRARIES to the profiler library. Unlike CUDA, it can't attach to a process that is already running.vLLM and SGLang, with HIP graphs. NVIDIA Dynamo-Triton and TensorRT-LLM are NVIDIA-only. See Inference.
AMD metrics are per device rather than per process, recommendations for ROCm are coming soon, and the GPU timeline is NVIDIA-only.
Profile your AMD fleet
Zymtrace is fully self-hosted. 5 minutes is all you need to get it up and running.