Member of Technical Staff - GPU Performance

The role

You will characterize and improve the performance of production AI inference across a heterogeneous accelerator fleet. You will build measurement systems and performance methodologies to understand hardware and implement changes across kernels, compilers, runtimes, and distributed execution.

What you will do

• Establish trustworthy performance baselines for production inference workloads
• Isolate bottlenecks across compute, memory, cache, kernel launch, synchronization, and scheduling
• Use profiling, hardware counters, and performance models to identify optimizations
• Improve execution through launch reduction, graph capture, and memory reuse
• Build automated benchmarks and regression detection
• Translate performance findings into production improvements
• Develop a repeatable methodology for evaluating and bringing up new accelerators

Required qualifications

• Strong software engineering fundamentals
• Experience working on performance-critical systems
• Experience profiling GPU or accelerator workloads and turning measurements into production improvements
• Strong understanding of GPU execution, memory hierarchies, asynchronous execution, synchronization, and performance tradeoffs
• Ability to design controlled experiments and communicate performance findings precisely
• Bachelor's degree in a relevant field, or an equivalent combination of education, training, and professional experience

Nice to have

• Experience with Nsight Systems, Nsight Compute, rocprof, Omniperf, VTune, Perfetto, or hardware performance counters
• Experience with CUDA, HIP, Triton, or another accelerator programming model
• Ability to prototype low-level optimizations
• Familiarity with roofline analysis, queueing behavior, occupancy, latency hiding, instruction throughput, and memory-bandwidth analysis
• Experience optimizing inference runtimes, execution graphs, memory allocators, kernel dispatch, batching, or model-serving systems
• Familiarity with collective communication, topology-aware execution, and compute-communication overlap across multiple accelerators
• Experience building performance laboratories, benchmark suites, regression infrastructure, or hardware bring-up methodology

About the company

Gimlet is building the first multi-silicon neocloud designed for fast, efficient AI inference. They combine large-scale compute infrastructure with an execution platform that partitions AI workloads.