Founding Kernel Engineer, GPU Kernels

Role overview Develop and qualify GPU kernels that improve real model workloads while preserving numerical correctness. The work spans framework dispatch, compilers, device code, and profiling across accelerator architectures; optimization choices should be supported by reproducible measurements and an end-to-end workload benefit.

Responsibilities - Find costly operators and execution gaps from representative workload traces. - Implement, select, fuse, or generate kernels for relevant operations such as attention, matrix math, routing, normalization, and data movement. - Trace framework operations through graph capture, compiler lowering, runtime dispatch, and device execution. - Profile launch overhead, occupancy, instruction mix, memory traffic, synchronization, stalls, and numerical error. - Compare vendor libraries, compiler paths, generated kernels, and custom implementations before taking on maintenance burden. - Build correctness tests, benchmark harnesses, architecture guards, fallbacks, and end-to-end performance evidence.

Requirements - Strong C++ and GPU programming fundamentals. - Experience with CUDA, ROCm/HIP, Triton, or a comparable accelerator programming environment. - Understanding of memory hierarchy, parallel decomposition, synchronization, precision, and numerical correctness. - Ability to use profilers, disassemblers, and benchmarks to substantiate performance claims. - Familiarity with framework dispatch and the compiler/runtime path around kernels. - Ability to maintain architecture-specific guards, correctness coverage, and portable fallbacks.

Nice to have Compiler IR or code generation; attention, GEMM, MoE, or communication kernels; collectives and multi-GPU execution; mixed-precision compute; or upstream open-source contributions.