GPU Kernel Engineer

Project

Work done for a product or program.

• Tune tile shapes, warp specialization, register allocation and shared-memory swizzle patterns to eliminate bank conflicts and maximize occupancy.
• Implement asynchronous multi-stage software pipelines using TMA or DMA descriptors to overlap global-memory loads with matrix-multiply issue.
• Analyse profiler traces and hardware counters to attribute stalls to memory bandwidth, instruction issue, synchronization or tail effects.
• Implement mixed-precision variants in FP16, BF16, FP8 and MXFP4 while maintaining accuracy targets agreed with the numerics team.
• Maintain autotuning configuration databases and heuristics that pick the best kernel variant per problem shape at runtime.
• Provide pre-silicon performance feedback to architects using RTL simulators, emulation and cycle-approximate models.

Function

Work that keeps the discipline running: methods, tools, suppliers and people.

• Write fused attention, GEMM and convolution kernels in CUDA, HIP, Triton, OpenCL or a vendor intrinsics layer targeting tensor cores or matrix engines.
• Benchmark each kernel against roofline and vendor library baselines such as cuBLAS, cuDNN, CUTLASS and rocBLAS and document the gap.
• Collaborate with compiler engineers so that hand-tuned kernels are callable as library primitives or reproducible as generated code.
• Write unit and numerical accuracy tests plus performance regressions for every kernel entering the library.

Project

1. Produce and hand over the role's key deliverables
   • Production kernel library releases covering GEMM, attention and normalization primitives
   • Autotuned kernel configuration database keyed by problem shape and data type
   • Kernel-level performance reports versus roofline and competitor libraries
   • Numerical accuracy test vectors and tolerance specifications per kernel
   • Pre-silicon kernel performance projections feeding architecture trade studies
2. Develop the softwareGeneral
   • Design to the hardware specification
   • Write and review the code
   • Test on models, emulation or silicon
3. Release and supportGeneral
   • Package and document releases
   • Fix customer-reported defects

Function

1. Maintain the platformGeneral
   • Keep build, test and release tools working
   • Track toolchain and operating system changes
2. Raise qualityGeneral
   • Run code reviews
   • Extend automated tests