Gimlet
Member of Technical Staff - Kernels
San Francisco, CA. Posted 8 months ago
The role
You will build and optimize low-level execution primitives to turn accelerator performance into production inference performance. You will work across accelerators with different execution models, memory hierarchies, and software stacks.
What you will do
- Build and optimize kernels that improve latency, throughput, and hardware utilization for production AI workloads
- Develop execution strategies that unlock performance across both established and emerging accelerator architectures
- Improve memory efficiency, scheduling behavior, and execution characteristics across the inference stack
- Partner with compiler, runtime, and distributed systems engineers to ensure end-to-end performance optimization
- Influence how heterogeneous hardware is deployed and utilized within the next generation of AI infrastructure
- Help establish performance engineering standards that shape the future of Gimlet's execution platform
Required qualifications
- Strong software engineering fundamentals
- Experience working on performance-critical systems close to hardware
- Comfort reasoning about low-level execution behavior, memory hierarchies, and performance tradeoffs
- Bachelor's degree in a relevant field, or an equivalent combination of education, training, and professional experience
Nice to have
- Experience with CUDA, Triton, CUTLASS, or other accelerator programming models
- Deep understanding of GPU execution models (warps/wavefronts, blocks, grids)
- Experience optimizing memory access patterns (coalescing, shared memory, cache behavior)
- Familiarity with occupancy, latency hiding, and instruction-level parallelism
- Experience using profiling and performance analysis tools
- Familiarity with multi-GPU or distributed execution
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 and maps each stage to the hardware best suited to run it.