TraceLens compatibility matrix

2026-07-22

3 min read time

This topic lists the hardware and software configurations for TraceLens. Only configurations that have been verified and tested should appear in the validated tables below.

ROCm and CUDA compatibility

TraceLens analyzes trace files rather than talking to the GPU or runtime directly, so it isn’t tied to a specific ROCm version. rocprofv3-based capture (JSON and pftrace) only requires a ROCm installation that provides rocprofv3 and, for .pftrace conversion, traceconv; any ROCm version that emits well-formed traces should work. If a particular ROCm build produces malformed traces (for example, missing required fields), TraceLens reports the problem and stops rather than silently producing misleading results.

The PyTorch report workflow works on any torch.profiler trace regardless of the GPU backend, and is well tested on both AMD ROCm and NVIDIA CUDA traces.

Hardware requirements

TraceLens runs on a CPU host. It parses trace files and never executes kernels or talks to a GPU, so no GPU is required to install or run it — you can analyze traces on a laptop or CI runner with no GPU present.

Target GPU for roofline analysis

The GPU details below describe the target GPU whose traces you analyze (the tracing environment), not a requirement for the machine running TraceLens. Roofline classification needs a GPU architecture specification that provides the peak FLOPS and peak bandwidth for that target GPU (for example, mi300.json). Architecture specifications are bundled under TraceLens/Agent/Analysis/utils/arch/; supply your own JSON with --gpu_arch_json_path for GPUs not listed.

Target GPU	Notes
AMD Instinct™ MI300X	Used for the roofline knee point (for example, FP16 ≈ 1300 TFLOPS / 5.3 TB/s ≈ 245 FLOPs/byte).

Software requirements

TraceLens requires the following software:

Component	Requirement	Notes
Python	3.6 or later (3.10 recommended)	python_requires>=3.6; the 3.10 toolchain is used for documentation and CI builds.
Operating system	N/A	TraceLens analysis and report generation are pure Python and run on any OS. Capturing ROCm-based traces requires a Linux ROCm environment; TraceLens only reads those traces afterward.
TraceLens package	0.1.0	Installed from github.com/AMD-AGI/TraceLens.

Python dependencies

These are installed automatically with the package:

Dependency	Purpose
pandas	Tabular analysis and report data frames.
openpyxl	Excel (.xlsx) report output.
tqdm	Progress bars for long-running analyses.
orjson	Fast JSON parsing of large traces.
tabulate	Text/Markdown table rendering.
matplotlib	Roofline and other plots.
xprof==2.20.1	JAX XPlane parsing (HLO sidecar generation; supports JAX 0.8+).
protobuf>=6.31.1,<7.0.0	Required by xprof’s grpcio-status dependency.
backports.strenum	StrEnum backport for Python < 3.11.
office365-rest-python-client, msal	Optional SharePoint/365 integrations.
traceconv	Optional; required only for .pftrace input. Resolved from PATH or downloaded automatically if not provided with --traceconv.

Supported trace formats

TraceLens supports the following trace formats:

Format	Producing tool
PyTorch Chrome trace (.json, .json.gz, .zip)	torch.profiler
JAX XPlane protobuf (.pb)	JAX profiler / xprof
rocprofv3 JSON (*_results.json)	AMD ROCm ROCprofiler-SDK
rocprofv3 pftrace / Perfetto-style	rocprofv3 --output-format pftrace