GPU Kernel Engineer

About Relace

Relace is building a globally distributed inference cloud. Customers send us requests, we run them on our own fleet of GPUs across the world, and we compete on price, speed and reliability against companies far larger than us. We have customers and revenue today, we’re venture backed, and the team is still small, so roles come with a lot of responsibility and room to grow.

We build our own stack. The routing layer is our own Rust service; the fleet runs on Kubernetes across multiple clouds. Performance here is the business: making the system ten percent faster makes the company roughly ten percent more money, and we can show you the number.

Some numbers: we process 800B+ tokens a day, we’ve grown 3x in the last three months, we are regularly one of the top providers by volume on OpenRouter, and we raised a $23M Series A led by Andreessen Horowitz, with Matrix Partners and Y Combinator. Lovable, Magic Patterns, Orchids, Codebuff, Create and Tempo Labs have used our technology.

The role

You would work at the bottom of the stack, writing and tuning the code that runs directly on the GPU. The job is making the hardware do more work per second and per dollar: writing fast GPU code, shrinking models so they’re cheaper to run without making them worse, and hunting down the many boring reasons a GPU sits idle.

What we need to see

- Strong C or C++ and genuine comfort at a low level: memory, concurrency, caches, hardware behaviour.
- A track record of making something significantly faster, where you can explain precisely why it was slow and what you changed. This matters more than anything else on this list.
- The habit of measuring. You profile before and after, and you’re suspicious of numbers that came from a sloppy test setup.

What we don’t require

GPU or CUDA experience. We’d rather have an excellent low-level systems engineer with no GPU background than an average CUDA programmer; learning to program a GPU on top of that takes a few months. Machine learning research experience isn’t needed either.

Nice to have

Operating system, hypervisor, driver, firmware or real-time systems work; performance tools you built for other engineers to use; compiler work; and, of course, actual GPU experience, whether at NVIDIA or AMD or through open-source work on the common inference projects.

A question we’ll ask

What’s the biggest speedup you personally delivered, what was the bottleneck, and how did you find it?

Working here

San Francisco, in person, five days a week. This is a senior role, with competitive compensation plus equity. Nine engineers and no go-to-market team means you own real surface area from the start, and the distance between your work and the company’s numbers is very short. Founded in 2023 by Preston Zhou and Eitan Borgnia.