Machine Learning Performance Engineer

You will work in a small and highly technical team operating at extremely high and low levels of abstraction to dramatically increase the performance of the firm's training and inference platforms. You won't just write CUDA code — you'll build a sophisticated optimising compiler for a variety of accelerated computing platforms. You'll need to be proficient in memory hierarchies, degrees of parallelism, microarchitectural nuances around instruction dispatch rates, and irregular bitwidth numerics. You'll also help design novel hardware architectures for accelerating models, mapping AI models from JAX graphs down to transistors.

Responsibilities

• Increase the performance of training and inference platforms
• Build and maintain an optimising compiler for accelerated computing platforms
• Work with memory hierarchies and degrees of parallelism
• Analyze microarchitectural nuances on instruction dispatch rates and irregular bitwidth numerics
• Assist in the design of novel hardware architectures for accelerating models
• Map AI models from JAX graphs to transistors

Requirements

• Experienced software engineer, likely in C/C++/Rust or a similar language
• Confident solving complex algorithmic problems
• Understanding of, and some experience targeting, GPUs, TPUs, or similar hardware
• Some experience working with data-flow oriented compilers for AI or scientific computing
• Writing Linux kernel modules or other experience in OS development

Benefits

• Onsite gym, sauna, and fitness classes at no charge
• Extensive medical benefits including an on-site doctor and therapist at no charge
• Breakfast and lunch provided daily
• Various supports for caregivers, including emergency dependent care
• 25 days paid holiday per year plus statutory holiday and paid sick days
• Generous pension contributions