Machine Learning Performance Engineer
Jane Street2 months ago
Responsibilities
- Optimize machine-learning model training performance at large scale.
- Optimize low-latency inference for real-time systems and high-throughput inference for research.
- Improve CUDA performance using a whole-systems approach covering storage, networking, host systems, GPUs, and memory hierarchy.
- Debug training-run performance end to end and evaluate whether system throughput represents useful goodput.
- Use GPU, networking, and distributed-training technologies to improve performance across GPU clusters.
Requirements
- Experience in low-level systems programming and optimization.
- Understanding of modern machine-learning techniques and toolsets.
- Systems knowledge and experience debugging training-run performance end to end.
- Low-level GPU knowledge of PTX, SASS, warps, cooperative groups, Tensor Cores, and GPU memory hierarchy.
- Debugging and optimization experience with CUDA GDB, Nsight Systems, and Nsight Compute.
- Knowledge of Triton, CUTLASS, CUB, Thrust, cuDNN, and cuBLAS.
- Understanding of CUDA graph launch, Tensor Core arithmetic, warp-level synchronization, and asynchronous memory loads.
- Background in Infiniband, RoCE, GPUDirect, PXN, rail optimization, and NVLink for linking GPU clusters.
- Understanding of collective algorithms for distributed GPU training in NCCL or MPI.
- Inventive problem-solving approach and willingness to question existing approaches and tools.
Tech Stack
Sass
Categories
About Jane Street
Jane Street is a global quantitative trading firm and liquidity provider that builds in-house software and research platforms to trade across asset classes. It makes markets and executes proprietary strategies on exchanges and electronic venues, serving institutional markets rather than individual investors. Founded in 2000 and headquartered in New York, it is privately held with offices in London, Hong Kong, Singapore, and Amsterdam.