GPU Kernel Performance Architect for Production AI

ATBF Labs; San Francisco (CA); Hybrid; USD 210,000 - 275,000; Full time

Benefits offered by this job

Health, dental, vision coverage

Equity

Job summary

ATBF Labs seeks a Member of Technical Staff to own performance at the metal, developing GPU kernels, memory movement, and numerics for low-precision safety. You will ship kernels that dramatically reduce latency and cost for production AI inference.

The role requires deep GPU architecture knowledge, CUDA/ROCm proficiency, and a track record of measurable speedups. You’ll collaborate with runtime and research teams to maximize hardware efficiency.

Qualifications

• Experience writing and optimizing GPU kernels (CUDA, Triton, CUTLASS).
• Profiling with Nsight/CUPTI to identify bottlenecks and optimize memory and compute.
• Experience with mixed-precision or quantization schemes.
• Familiarity with multi-GPU/multi-node communication (NVLink, InfiniBand, RoCE).
• Strong understanding of GPU architecture, memory hierarchy, and parallel programming.
• Experience with ML compilers and model inference stacks.

Responsibilities

• Write and optimize GPU kernels (CUDA, Triton, CUTLASS) for attention, GEMMs, and quantized paths.
• Profile with Nsight / CUPTI to find kernel- and memory-level bottlenecks, then close them.
• Implement mixed-precision and quantization schemes (FP8, INT4, and beyond) with measured quality bounds.
• Optimize communication for multi-GPU and multi-node serving over NVLink and RDMA (InfiniBand, RoCE).
• Co-design model execution with the runtime and research teams for hardware efficiency.
• Debug numerical instabilities that only appear at scale, on a fraction of requests.
• Deep understanding of GPU architecture, parallel programming, and the memory hierarchy.
• Proficiency in CUDA (or ROCm) and a GPU profiler (Nsight, nvprof, CUPTI).
• Experience with performance-critical model execution and PyTorch internals.
• A track record of measurable speedups, not just refactors.
• Implement a fused attention kernel for a novel variant and beat the reference by a measurable margin.
• Ship an FP8 path and run experiments to find the optimal speed/quality tradeoff.
• Optimize RDMA communication patterns to remove a stall in multi-node decode.

Skills

CUDA

Triton

CUTLASS

Nsight

nvprof

CUPTI

PyTorch

GPU architecture

parallel programming

memory hierarchy

ROCm

torch.compile

XLA

TVM

Tools

Nsight

nvprof

CUPTI

torch.compile

Triton

XLA

TVM

Job description

ATBF Labs seeks a Member of Technical Staff to own performance at the metal, developing GPU kernels, memory movement, and numerics for low-precision safety. You will ship kernels that dramatically reduce latency and cost for production AI inference.

The role requires deep GPU architecture knowledge, CUDA/ROCm proficiency, and a track record of measurable speedups. You’ll collaborate with runtime and research teams to maximize hardware efficiency.