Remote Kernel Performance Engineer – AI Inference

Inferact; San Francisco (CA); Remote; USD 200,000 - 400,000; Full time

Benefits offered by this job

Health benefits including vision, and

401(k) company match

Job summary

Inferact is seeking a Senior Performance Engineer to squeeze every FLOP out of modern accelerators. You will write CUDA kernels and low-level optimizations to make vLLM the fastest inference engine.

Your code runs on hundreds of accelerator types, and you’ll work with hardware teams to extract maximum performance from every generation of hardware. You will collaborate across NVIDIA, AMD, TPU, and Intel ecosystems, profiling and tuning to achieve best-in-class speedups while supporting ongoing ML

Qualifications

• Bachelor's degree or equivalent experience in computer science, engineering, or similar.
• Deep experience writing CUDA kernels or equivalent (CuTeDSL, Triton, TileLang, Pallas).
• Strong understanding of GPU architecture: memory hierarchy, warp scheduling, tiling, tensor cores.
• Proficiency in C++ and Python with demonstrated ability to write high-performance code.
• Experience with profiling tools (Nsight, rocprof) and performance optimization methodologies.
• Obsession with benchmarks and squeezing every percentage point of speedup.

Responsibilities

• Write high-performance CUDA kernels and low-level optimizations.
• Develop portable kernels across diverse accelerator types (NVIDIA, AMD, TPU, Intel).
• Collaborate with hardware teams to ensure peak performance on new chips.
• Profile, benchmark, and iterate to maximize speedups across generations.

Skills

CUDA kernel development

C++ and Python

GPU architecture and optimization

Profiling and performance tuning

Benchmarking focus

Remote collaboration with hardware/ops

ML kernel optimization

Quantization techniques

Multiple accelerators knowledge

Compiler technologies experience

Education

Bachelor's degree or equivalent in CS/Engineering

Job description

Inferact is seeking a Senior Performance Engineer to squeeze every FLOP out of modern accelerators. You will write CUDA kernels and low-level optimizations to make vLLM the fastest inference engine.

Your code runs on hundreds of accelerator types, and you’ll work with hardware teams to extract maximum performance from every generation of hardware. You will collaborate across NVIDIA, AMD, TPU, and Intel ecosystems, profiling and tuning to achieve best-in-class speedups while supporting ongoing ML