Open for remote contract work — GPU/CPU performance work, kernel optimization and benchmarking.
Performance engineer. I make code go fast on multicore CPUs and GPUs, and I measure what the hardware can actually do rather than what the spec sheet claims.
Author and maintainer of clpeak — a synthetic micro-benchmark that exposes peak compute, bandwidth, and matrix throughput across OpenCL, CUDA, ROCm/HIP, Metal, Vulkan, oneAPI/SYCL, and native CPU. Hand-rolled MMA kernels (WMMA, MFMA, simdgroup_matrix, coopmat, AMX/SMMLA) alongside vendor GEMM paths (cuBLASLt, rocBLAS/hipBLASLt, MPS, oneMKL), down to fp8 and MXFP4/NVFP4.