SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
Relace is building a globally distributed inference cloud that processes 800B+ tokens per day. The company competes on price, speed, and reliability against much larger competitors, with customers and revenue today. Backed by Andreessen Horowitz (Series A, $23M), Matrix Partners, and Y Combinator, the team remains small with significant room for growth and responsibility.
In this role, you will work at the bottom of the stack, writing and tuning code that runs directly on GPUs. Your focus will be on making hardware perform more work per second and per dollar, optimizing model sizes for cost-efficiency without degrading quality, and identifying and eliminating GPU idle time. The company emphasizes that performance directly impacts revenue: a 10% system speedup translates to roughly 10% more company revenue.
The infrastructure runs on Kubernetes across multiple clouds, and you'll be part of a team that regularly ranks among the top providers by volume on OpenRouter.
Key responsibilities include:
- Writing and optimizing GPU kernel code for maximum performance
- Profiling and tuning system performance across the inference pipeline
- Identifying bottlenecks and implementing solutions to reduce GPU idle time
- Collaborating with the team to balance performance, cost, and reliability
- Measuring and validating improvements rigorously before and after changes
The company is explicit that they value excellent low-level systems engineers over average CUDA programmers; GPU programming can be learned in a few months, but deep systems thinking cannot.
Working arrangement: San Francisco, in-person, five days per week. This is a senior role with competitive compensation plus equity.
Requirements:
- Approximately five or more years of low-level performance work in domains where system speed was critical: operating systems, drivers, databases, compilers, trading systems, or inference engines
- Strong C or C++ skills with genuine comfort at the low level: memory management, concurrency, caches, and hardware behavior
- Demonstrated track record of making something significantly faster, with ability to explain precisely why it was slow and what changes were made (this is weighted most heavily)
- Disciplined approach to measurement: profiling before and after changes, skepticism of results from sloppy test setups
Nice to have:
- Operating system, hypervisor, driver, firmware, or real-time systems experience
- Performance tools built for other engineers
- Compiler work experience
- Actual GPU experience (NVIDIA, AMD, or open-source inference projects)
