Performance Engineer, GPU

About the role

Anthropic is looking for a GPU Performance Engineer to design and implement the core systems that power its large language models. You will work at the intersection of hardware and software, driving GPU utilization and efficiency at massive scale.

Key responsibilities

• Architect and develop custom GPU kernels and low‑level optimizations for tensor cores.
• Design distributed communication strategies across thousands of GPUs, ensuring synchronized execution.
• Build performance‑modeling tools to predict and improve GPU utilization.
• Collaborate with researchers to co‑design attention mechanisms and quantization techniques for next‑generation hardware.
• Maintain and scale large‑scale training and inference infrastructure with fault‑tolerance and orchestration.

Required profile

• Proven track record of delivering measurable GPU performance gains in production ML systems.
• Strong problem‑solving mindset, comfortable navigating hardware interfaces to high‑level ML frameworks.
• Passion for AI safety and societal impact of technology.
• Ability to thrive in ambiguous, fast‑moving environments and work collaboratively.

Required skills

• GPU programming and optimization: CUDA, Triton, CUTLASS, Flash Attention, tensor‑core optimization.
• ML compilers and frameworks: PyTorch, JAX internals, torch.compile, XLA, custom operators.
• Performance engineering: kernel fusion, memory‑bandwidth optimization, profiling with Nsight.
• Distributed systems: NCCL, NVLink, collective communication, model parallelism.
• Low‑precision techniques: INT8, FP8 quantization, mixed‑precision.
• Production infrastructure: large‑scale training pipelines, fault tolerance, cluster orchestration.