GPU Performance Engineer: Scale Inference & Training

Anthropic; York and North Yorkshire; On-site; GBP 90,000 - 140,000; Full time

Benefits offered by this job

Comprehensive health insurance

Fertility benefits

22 weeks parental leave

Flexible paid time off

Job summary

Anthropic seeks a GPU Performance Engineer to architect and implement foundational systems powering Claude, driving GPU utilization and optimization at unprecedented scale. You will work at the hardware-software boundary, from custom kernel development to distributed multi-node clusters.

Responsibilities include end-to-end optimization of training and inference pipelines, co-design of attention mechanisms for future architectures, and collaboration with hardware vendors to shape accelerator

Qualifications

• Track record delivering GPU performance improvements in production ML systems.
• Experience optimizing end-to-end training and inference pipelines at scale.
• Strong understanding of modern ML frameworks and low-level GPU hardware.

Responsibilities

• Architect and implement foundational GPU performance systems for Claude.
• Maximize GPU utilization and inference efficiency at scale.
• Develop kernel-level optimizations and distributed architectures across thousands of GPUs.
• Co-design attention mechanisms for next-gen hardware architectures.
• Collaborate with hardware vendors to influence accelerator capabilities.

Skills

GPU performance optimization

Collaborative problem-solving

Hardware-software co-design

Education

Bachelor's degree or equivalent experience

Tools

CUDA

Triton

CUTLASS

Flash Attention

Tensor cores

PyTorch/JAX internals

torch.compile

XLA

NCCL

NVLink

Job description

Anthropic seeks a GPU Performance Engineer to architect and implement foundational systems powering Claude, driving GPU utilization and optimization at unprecedented scale. You will work at the hardware-software boundary, from custom kernel development to distributed multi-node clusters.

Responsibilities include end-to-end optimization of training and inference pipelines, co-design of attention mechanisms for future architectures, and collaboration with hardware vendors to shape accelerator