Graphsignal

Graphsignal is a GPU profiler for inference workloads, built so that an AI agent can tune performance as well as a human engineer. It measures where time actually goes during generation and serves the results in a form a coding agent can read, act on by changing a flag or a line of code, and then measure again. It is aimed at inference engineers and teams tuning model serving on their own hardware. The profiler reports cumulative time per kernel, per CUDA graph and per memory or synchronization operation through CUPTI and ROCm, alongside GPU telemetry, engine metrics and error capture, and it covers NVIDIA and AMD accelerators with PyTorch, vLLM, SGLang and TensorRT. A vendorable C++ header provides GPU probes that instrument application code and kernels with lock free recording, stay inert when the profiler is not attached, and surface probe values next to the built in profiles. Everything measured is served from a single local JSON endpoint, and a companion terminal view renders the same data live. Signals can optionally upload so that agents read production behavior back through an API; without an API key nothing is sent. A free plan costs $0 and includes 100 profiled hours a month, telemetry, observability dashboards and a single user, where a profiled hour means one GPU or accelerator profiled for one hour. Pro is $0.08 per profiled hour and adds continuous production profiling, unlimited users and thirty days of data retention. Enterprise is priced on request and adds an on premise option and volume discounts. Subscriptions bill at the start of each monthly or annual cycle and run to the end of the paid period after cancellation.