Machine Learning Scientist (Quantisation) in UNITED KINGDOM

Fractile is building the hardware and software needed to run the world’s largest AI models 100x faster than we can run them today, at a fraction of the cost. Doing this successfully requires innovating across the stack, from custom silicon to the software that runs on it. Building hardware for AI acceleration is a unique challenge — workloads can change overnight as new models are released and new architectures invented, while hardware can take years to bring to market. Fractile is a team of incredibly talented engineers and scientists who are committed to tackling this challenge and completely transforming the performance of the world’s leading AI models. 🚀🚀🚀

We are looking for a machine learning scientist with experience in model quantisation to join us at Fractile. The past few years have shown the significance of leveraging lower-precision numerics as part of any performance optimisation for model inference, and for the largest frontier models, the best quantisation recipes are still active areas of research. In this role at Fractile, you would be expected to engage with that research and drive our own experimentation with cutting edge quantisation and numerics. The luxury of building custom silicon is that we are not constrained to a small set of conventional number formats and operators, and there is an opportunity to have extraordinary impact on performance through the close co-design of numerics, quantisation schemes, and hardware.

NB: Applicants who don’t have direct prior experience with model quantisation research but do have very strong ML backgrounds including LLM experience are still encouraged to apply!

In this role, you will:

• Stay up to date with bleeding-edge work on model quantisation and number formats leveraged for efficient frontier model inference
• Contribute to Fractile’s R&D on quantisation and numerics, and work with Fractile’s ML compiler and hardware micro-architecture teams to drive product direction and development as pertains to numerics and supported number formats
• Engage with the broader research community and Fractile’s commercial partners to understand the trends in numerical formats and quantisation methods used in frontier model inference

You have:

• Strong understanding of the state of the art in post-training quantisation and quantisation-aware training methods for neural network inference
• Good understanding of LLM inference performance optimisation methods, including state-space models, kv-cache compression, dynamic batching, multi-user inference settings, tensor and pipeline parallelism, etc.
• A desire to learn new things and work at the challenging intersection of cutting edge research and product specification, and to work with both software and hardware teams within the business.

You may also have:

• Experience with hardware design, including processor microarchitecture
• Previous experience in a startup or small team environment

Location:

UK. You must be able to commute to our offices in London (Farringdon) or Bristol (city centre) for at least two days per week. Fractile can sponsor visas and can help with relocation.