GPU Systems Engineer — Distributed Training & Inference

TensorScale AI; San Francisco (CA); On-site; USD 180,000 - 240,000; Full time

Job summary

TensorScale AI in San Francisco seeks a hardware‑aware software engineer to optimize GPU systems for training and inference across image, video, and world‑model workloads. You will push performance from kernel code to distributed engines, profiling bottlenecks and implementing practical improvements that scale.

Responsibilities include CUDA / Triton optimizations, designing efficient distributed inference and training pipelines, and owning communication performance across GPUs and nodes with

Qualifications

• 3+ years in deep learning inference/training systems, distributed systems, or high-performance computing.
• Proficiency in CUDA, with hands-on GPU profiling (Nsight).
• Strong systems instincts across GPU architecture, parallel programming, and compute kernels.
• Experience with distributed multi-GPU debugging and optimization.
• Familiarity with PyTorch and performance-critical model execution.
• Comfort working onsite with the team in the Bay Area.

Responsibilities

• Optimize system and GPU performance of training and inference for image, video, and world-model workloads.
• Profile and remove bottlenecks at kernel, memory, system, and cluster levels (Nsight and related tooling).
• Implement low-level optimizations in CUDA / Triton.
• Design and improve distributed inference / training engines for diffusion models.
• Own communication performance across GPUs and nodes: NCCL, RDMA / InfiniBand / RoCE, and disaggregated serving.
• Build benchmarking and regression harnesses so performance gains stick in production.
• Hardware-aware kernel, runtime, and model co-design.

Skills

CUDA proficiency

Nsight profiling

Distributed systems

PyTorch

Multi-GPU debugging

Performance optimization

Tools

Nsight

NCCL

InfiniBand

RoCE

TensorRT

Job description

TensorScale AI in San Francisco seeks a hardware‑aware software engineer to optimize GPU systems for training and inference across image, video, and world‑model workloads. You will push performance from kernel code to distributed engines, profiling bottlenecks and implementing practical improvements that scale.

Responsibilities include CUDA / Triton optimizations, designing efficient distributed inference and training pipelines, and owning communication performance across GPUs and nodes with