TPU Kernel Engineer

About the role

Anthropic is looking for a TPU Kernel Engineer to improve performance across a variety of machine‑learning workloads, from research experiments to large‑scale training and inference. You will design, implement, and optimise low‑level kernels for Google’s Tensor Processing Units, collaborating closely with researchers to understand how model changes affect system behaviour.

Key responsibilities

• Identify and resolve performance bottlenecks in ML systems on TPUs, GPUs and other accelerators.
• Design, implement and optimise custom kernels for high‑throughput, low‑latency sampling of large language models.
• Adapt existing models for low‑precision inference and build quantitative performance models.
• Develop custom collective communication algorithms and debug kernel performance at the assembly level.
• Provide feedback to researchers on the performance impact of model modifications.

Required profile

• Proven experience solving large‑scale systems problems and low‑level optimisation.
• Results‑oriented mindset with a bias toward flexibility and impact.
• Willingness to pick up tasks outside the strict job description and enjoy pair programming.
• Interest in machine‑learning research and its societal implications.

Required skills

• Deep knowledge of TPUs, GPUs or other ML accelerators.
• Kernel development and optimisation for high‑performance ML workloads.
• Understanding of computer architecture and accelerator internals.
• Familiarity with ML framework internals and transformer‑based language models.
• Experience with low‑precision inference, collective communication, and assembly‑level debugging.

What we offer

• Competitive annual compensation ranging from $280,000 to $850,000 USD.
• Hybrid work model with at least 25 % onsite presence in the San Francisco office.
• Visa sponsorship where possible, with dedicated immigration support.