Systems Performance Engineer, AI Infrastructure

CommonAI CIC is a non-profit organisation built around collaborative engineering and is backed by a mix of UK Government and private funding in order to design, build and deploy innovative AI systems.

Our Scaling Inference Lab (SIL) is developing simulation grounded in real measurements to help the UK invest in promising AI infrastructure. We connect proposed hardware designs to realistic workloads, helping investors and engineers understand their potential, their limitations and the next experiment worth doing.

Design and run experiments that reveal how AI infrastructure behaves, what limits its performance and how reliably that behaviour can be predicted. You will build measurement tools and turn observations of real hardware and software into evidence for SIL's simulation platform and the investment decisions it supports.

Your work will range from targeted experiments on individual components to measurements of complete LLM serving systems. Working with workload, simulation and uncertainty specialists, you will investigate bottlenecks, explain performance variation and identify the meaningful effects.

We welcome deep expertise in one or more parts of the hardware and software stack. You will work with colleagues who bring complementary strengths, with scope to broaden your expertise as the team grows.

• Design and run reproducible experiments, from individual hardware or software components to complete LLM serving systems, covering relevant workloads, configurations and operating conditions
• Develop reusable measurement tools and collect traces, profiles and counters to characterise latency, throughput, memory behaviour and, where reliable measurement is possible, power and energy
• Investigate bottlenecks, performance variation and unexpected behaviour across the layers your expertise covers, working with colleagues to extend the investigation through the stack
• Take responsibility for reproducible workload details, software versions, configuration and measurement methods so results can be reproduced, compared and used to calibrate simulator models
• Work with simulation and uncertainty specialists to choose measurements that resolve important unknowns and undertake curation of calibration and validation data
• Help define the controls needed for experiments and use the resulting cluster-control interfaces, coordinating access and changes with colleagues responsible for the hardware

An example early project would be to work with workload colleagues to establish a reproducible reference experiment and measurement dataset, investigating how performance changes under a selected configuration or load change.

• Experience investigating the performance of real computing systems and designing experiments that support rigorous conclusions
• Depth in one or more relevant areas: accelerator profiling, inference runtimes, memory systems, distributed/HPC workloads, networking, or power and energy measurement
• The ability to build maintainable automation and analysis tools in a suitable software language, and to learn unfamiliar profiling tools
• Knowledge and care of experimental controls, repeatability, confounding effects and the limits of a measurement
• Some experience using AI coding agents, alongside sound debugging and review habits

LLM serving experience and familiarity with tools such as Nsight, PyTorch Profiler or system/network profilers would be useful. These are examples, not a requirement to know every tool. Strong expertise in one part of the stack is welcome; we can shape the role and recruit complementary strengths around it.

• Help build UK AI capability and improve how investment and compute serve the public good
• Build a network spanning UK…