Comparison with Other Profilers

Feature comparison

Feature	rocm-trace-lite	rocprofiler-sdk	roctracer + RPD
Dependencies	libhsa-runtime64 + libsqlite3	ROCm 6.0+ full stack	roctracer + RPD
GPU kernel timing	HSA signal injection	HW counters + callbacks	roctracer callbacks
HIP API tracing	No	Yes	Yes
HW performance counters	No	Yes	No
PC sampling	No	Yes	No
roctx markers	Built-in shim	libroctx64	libroctx64
Multi-GPU	Automatic per-process merge	Manual	Manual
Output format	SQLite .db + Perfetto JSON	Various	RPD SQLite
Overhead	Low (signal pool, single worker)	Medium-High	Medium
New HW bring-up	Works immediately (HSA only)	Requires rocprofiler support	Requires roctracer support

When to use rocm-trace-lite

• Kernel profiling on new hardware where rocprofiler is not yet available

• Lightweight CI regression testing of kernel performance

• AI framework teams who need clean kernel profiling without heavy dependencies

• Quick distributed profiling (TP>1) with automatic merge

When to use rocprofiler-sdk instead

• You need HW performance counters (cache hit rates, occupancy, etc.)

• You need PC sampling for hotspot analysis

• You need HIP API tracing (API call timestamps, arguments)

• You need the full ROCm profiling ecosystem

Relationship to RPD

rocm-trace-lite is a standalone alternative, not a fork of ROCm/rocmProfileData:

• Written from scratch (no RPD code dependency)

• Outputs RPD-compatible SQLite schema (interoperable with RPD tools)

• Can coexist with original RPD on the same system