Introducing G-Watch#

GPU kernel optimization has long been a craft reserved for experts staring at profiler dumps. G-Watch is built to change that — it gives both engineers and coding agents the instruments to see inside a kernel and act on what they find.

G-Watch is a toolbox for agentic GPU/TPU kernel optimization, organized around two pillars.

Profiling#

Rich, runtime hardware observation across NVIDIA and AMD GPUs (and Google TPUs):

• Range profiling — hardware performance counters per launch, to pin the macro bottleneck (memory- vs compute-bound, pipe pressure, occupancy).
• PC sampling — instruction-level hotspots and the dominant warp-stall reasons, mapped to source and SASS lines.
• Intra-kernel tracing — the in-kernel phase timeline, exposing pipeline bubbles and synchronization overhead. The only tool with this temporal view.

Program analysis#

Tools for inspecting compiler-generated GPU/TPU binaries — demangle kernels, walk decoded SASS, map PC→source line, and dump the control-flow graph. This binary analysis foundation also powers secondary-development tasks such as register analysis and binary instrumentation.

Built for agents#

On top of the raw capabilities, G-Watch ships agent skills that drive a disciplined, reasoning-driven auto-optimization loop: profile → hypothesize → apply one minimal change → verify correctness, performance, and hardware behavior → repeat. Every iteration is a reasoned response to the previous one's on-disk evidence — no speculative or shotgun attempts.

The whole stack is DSL-agnostic: CUDA C++, Triton, CuTe DSL, TileLang, and HIP on AMD all work through the same workflow.

Get started#

pip install gwatch
npx skills add mars-compute-ai/G-Watch -a claude-code -a codex -a gemini-cli

Then head to the docs to run your first diagnosis.