Hey Ph! Aditya here, co-founder of nCompass.
My team and I have spent years accelerating GPU kernels. It’s complicated and time-consuming, mainly because we spent most of our time identifying which kernels to improve and then identifying the bottlenecks in our new kernel. Using the dev tool we’re launching today, we implemented a Hopper GEMM kernel that outperformed NVIDIA’s CUTLASS GEMMs by 3% within a day – it took us months.
Today’s issue is: If you profile a system like VLLM, you have to copy a large trace file to your local machine just for viewing. Then you spend hours figuring out which GPU cores to target. Then you profile the kernel, spending hours or days digging through .ncu traces that are massive data dumps. Then you identify your obstacles and make a plan. Running this loop until you have a performant kernel can take weeks, even months before you have a performant kernel.
This is all about to change. Coding agents have mastered writing GPU kernels once they know what to target. What is missing is to identify these goals.
We created the tool we wish we had. With our VSCode extension, you can view symbols natively – no more copying files. After our MCP is integrated, only the “@” cloud code / cursor has an execution trace and our agent takes over: our agent analyzes the trace and provides you with a preferred list of GPU kernels to target. Using NCU, profile your kernel source code, “@” .ncu-rep into your favorite coding agent, and the nCompass agent works with your coding agent to produce the high-performance kernel you’ve been wanting to build!
If you are an experienced GPU engineer. – It removes busyness. You stay in control, check the agent’s results, and only do what makes sense.
If you are learning GPU optimization. – Now you have an expert pair programmer you can always consult with. This is a great way to learn how to optimize GPU code.
We’d love for you to try it out and let us know how it works for you. We’re building fast and your feedback directly shapes what ships next.
Try it → /quick…
Complete feature list →