Research Engineer
How fast should an LLM building block run on our hardware?
About the Role
How fast should an LLM building block run on our hardware? How fast does it actually run, and what would make it faster?
As a Working Student, LLM Inference Performance, you'll help answer these questions by working directly with our hardware and software engineering teams.
You'll write and optimize low-level kernels, benchmark performance on real silicon, and develop profiling tools that help engineers understand and improve LLM inference performance.
You'll gain hands-on experience with our Rust programming stack and Python SDK, use AI coding agents as part of your development workflow, and collaborate closely with engineers working on compilers, runtimes, and kernel optimization.
This is an opportunity to apply your academic knowledge to real-world AI hardware challenges while taking ownership of meaningful technical projects.
What You'll Do
Understand our hardware: Learn how Tensordyne silicon processes data, manages memory, and synchronizes operations, and how key LLM components such as attention, MLP, and Mixture-of-Experts (MoE) routing map onto the architecture..
Develop and optimize kernels: Implement, test, and tune low-level kernels to improve the performance of LLM inference workloads..
Analyze performance: Build analytical and roofline performance models, compare predictions against simulator and real-silicon measurements, and investigate performance bottlenecks..
Build benchmarking and profiling tools: Develop benchmark harnesses, profiling scripts, and trace visualizations that make performance easier to measure, understand, and improve..
Work with AI coding agents: Explore workflows where AI agents assist with kernel development, optimization, and benchmarking, and identify the tools and context they need to work effectively..
Collaborate across engineering teams: Work closely with compiler, runtime, and kernel engineers to translate performance insights into practical improvements..
Example project: Develop a benchmarking and profiling harness for an LLM building block, such as decode attention, at two abstraction levels. Compare measured performance against theoretical estimates and make the tooling accessible to AI coding agents.
What We're Looking For
Currently enrolled in a Master's degree program in Computer Science, Electrical Engineering, Mathematics, or a related technical field..
Strong programming skills in Python and at least one of C, C++, or Rust..
Understanding of computer architecture, memory systems, bandwidth, and parallel computing through academic coursework or personal projects..
Interest in low-level programming, hardware performance, and optimization..
Practical experience using AI coding agents in development workflows, with the ability to evaluate and improve their output..
Posted November 2, 2026