Sim replaces libcuda.so, so your binary runs unchanged. It loads the compiled kernel into a SASS interpreter, runs it on CPU threads with a timing model, and reports the estimated time on each GPU and where each instruction stalls. It runs on a laptop or in CI.
The kernel page shows the SASS produced for each candidate, with the stall percentage for each instruction, register use, occupancy and the estimated time on each target GPU.
The agent edits the CUDA or Triton source, recompiles, reads the new SASS and simulates it. The best version is tested on a real GPU before a pull request is opened.
The kernel is compiled exactly as it would be for the GPU. Sim then runs the SASS on CPU threads and shows how long that took next to the estimated GPU time.
Occupancy from the CUDA occupancy-calculator rules for the chosen architecture, achievable bandwidth from Little's law, and the bound from a roofline. Sim goes further by interpreting the SASS; this shows the mechanics.
Static analysis takes milliseconds, sampled simulation takes seconds and cycle-level simulation takes hours. Use the fast ones to narrow the candidates down and the slow one to choose between the last few.
Each run reports how closely its ranking matches real hardware, and the top candidates are tested on a real GPU before anything ships.
Runs on GitHub Actions or any CPU runner, so a build can fail when a kernel gets slower on a GPU you don't own.
CUDA C++, Triton, CUTLASS, PyTorch through a driver shim · KernelBench and TritonBench · GitHub Actions · targets: H100, H200, B200, L40S, RTX 4090 and 5090, Jetson Orin and Thor.
We look at what the compiler actually produced: register use, spills, occupancy, shared-memory bank conflicts and stall reasons in the SASS. We don't write SASS by hand; ptxas does. We influence it with PTX intrinsics, launch bounds and compiler flags, then check the output.
Accurate enough to rank candidates, which is what it's for. Each run reports its rank correlation with real hardware, and nothing ships without a test on a real GPU.
Not yet. Sim only models NVIDIA architectures for now, so we tune AMD kernels on real hardware.
Send us the model, the hardware and your target, whether that's tokens per second, latency, memory or cost. In a 30-minute call we'll tell you whether we think it's doable and roughly what it would take.
