GPU Kernel Engineer

Relace•San Francisco, CA

2d•Onsite

About The Position

Relace is building a globally distributed inference cloud, running customer requests on its own GPU fleet around the world. The company builds its own stack, with the fleet running on Kubernetes across multiple clouds, and focuses on price, speed, and reliability. The role works at the bottom of the stack, writing and tuning code that runs directly on GPUs to increase throughput per second and per dollar, reduce model costs without degrading quality, and identify causes of GPU idleness. GPU or CUDA experience is not required; Relace values strong low-level systems engineering and performance instincts. This is a senior role based in San Francisco, in person five days a week, with competitive compensation plus equity.

Requirements

At least five years of low-level performance work where system speed mattered, such as operating systems, drivers, databases, compilers, trading systems, or inference engines.
Strong C or C++ skills and low-level understanding of memory, concurrency, caches, and hardware behavior.
A track record of significantly improving performance and being able to explain the cause of the slowdown and the changes made.
A habit of measuring and profiling before and after changes, and skepticism toward results from poor test setups.

Nice To Haves

Experience in operating systems, hypervisors, drivers, firmware, or real-time systems.
Experience building performance tools for other engineers.
Compiler experience.
GPU experience in industry or open source.

Responsibilities

Write and tune code that runs directly on GPUs.
Increase work completed per second and per dollar of GPU hardware.
Reduce model size and operating cost without degrading model quality.
Identify and resolve causes of GPU idleness.
Measure and profile performance before and after changes, using reliable test setups.