The posting
CommonAI CIC is a non-profit membership organisation, founded on a belief in collaborative engineering for the safe and responsible development of foundational AI technologies. A place where AI startups, enterprises large and small, public sector bodies and academia can share resources and knowledge, to codevelop and grow businesses, fast.
We are led by experienced founders, investors and engineers who believe that collaborative engineering drives faster AI innovation and are backed by a mix of UK Government and private funding in order to design, build and deploy innovative AI systems.
The Opportunity We are seeking a GPU Kernel and Performance Engineer to make our research models run fast on real hardware, primarily NVIDIA GPUs. Our research team works on reducing the complexity of large language models while retaining their performance. Some of this work informs our analyses of the trade-offs between training and inference costs when scaling inference workloads. The work runs on the UK's national AI supercomputers and on our own compute nodes.
You will join a small research team and turn new model designs into efficient GPU code, then measure how much time and memory traffic they cost on real hardware. Our training runs on NVIDIA GPUs, and we also expect to work with other AI accelerators, so there is scope to learn new chips and programming models as the work develops. We care more about how well you understand hardware and memory than about ML research credentials.
What You'll Do
- Write and optimise CUDA and Triton kernels on NVIDIA GPUs for sparse and structured model components, including block-sparse matrix operations and attention variants.
- Adapt and extend open-source kernels, such as FlashAttention-style and block-sparse attention, to new model designs.
- Profile training and inference to find where time and memory traffic go, using tools such as NVIDIA Nsight Systems and Nsight Compute.
- Optimise for the GH200's memory hierarchy on Isambard-AI and for our own GPU nodes.
- Build tools that measure data movement per generated token, and check analytic cost models against hardware measurements.
- Test every kernel against reference implementations before it runs at scale.
- Work with our training infrastructure engineer to raise training throughput, and with researchers to make new designs fast enough to test.
- Evaluate how our models map onto other AI accelerators, such as TPUs and dataflow, wafer-scale and compute-in-memory architectures, and port key components to them where it is useful.
- Contribute improvements back to open-source projects where possible.
Requirements
- Proven experience writing and optimising GPU kernels in CUDA or Triton, ideally on recent NVIDIA architectures.
- A strong understanding of GPU architecture and memory hierarchies (e.g. shared memory, HBM bandwidth, memory coalescing, occupancy).
- Hands-on experience profiling and optimising performance with tools such as NVIDIA Nsight Compute or Nsight Systems.
- Strong Python and C/C++.
- A rigorous approach to testing numerical code (e.g. comparison against reference implementations, mixed-precision error).
- Openness to learning new hardware architectures and programming models, often from limited documentation.
- A strong enthusiasm for using LLMs, coding assistants, and agents as core tools in your own software development process.
We also value:
- Experience with attention kernels (e.g. FlashAttention) or sparse kernels (e.g. block sparsity, mixture-of-experts routing).
- Experience with NVIDIA Hopper features (e.g. TMA, warpgroup matrix instructions) or libraries such as CUTLASS and CuTe.
- Experience with multi-GPU communication (e.g. NCCL, NVLink).
- An understanding of LLM inference performance (e.g. KV caching, memory-bound decoding, quantisation).
- Experience with other AI accelerators (e.g. AMD GPUs, TPUs with JAX and Pallas, or dataflow, wafer-scale and compute-in-memory architectures).
- Experience with compiler stacks for AI hardware (e.g. MLIR, XLA, TVM).
- A background in computer architecture, HPC, graphics or game engine programming.
- Experience with hardware/software co-design, or with hardware performance modelling and simulation.
- Contributions to performance-critical open-source projects (e.g. PyTorch, Triton, JAX, vLLM).
Benefits
- A collaborative and supportive work environment
- The opportunity to have a high impact in a growing organisation
- Competitive salary package and pension
- Professional development opportunities
- Networking opportunities with influential people from across the tech sector, financial services, and academia
- A vibrant office environment located a few minutes' walk away from Cambridge train station
CommonAI CIC is an equal opportunity employer and is committed to creating an inclusive and diverse workplace.
