Hiring nowAI compilers, kernels and runtime systems

GPU Kernel Engineer

Also called CUDA Engineer, AI · CUDA/Triton Engineer · Performance Engineer

A GPU Kernel Engineer writes and tunes the low-level code that runs model operations on accelerators. Faster, correct kernels cut compute cost and latency, and they decide how far existing hardware can be pushed.

What they do

The work starts with a model operation that runs too slowly or uses too much memory. The engineer profiles it, finds the bottleneck, and decides whether the limit is compute, memory bandwidth or launch overhead. They then write or rewrite the kernel in CUDA or Triton, adjusting tiling, data layout, shared memory use and scheduling. Every change is tested against a reference for numerical accuracy across many tensor shapes, not just one. They read compiler output and intermediate representations when the result looks odd. They also work with researchers who want new operations, and with runtime and compiler engineers who need kernels that fit the wider system.

What a strong hire delivers

In the first months, a strong hire gives you measured gains on the operations that cost the most. Expect tested kernel implementations with clear profiling results, showing speedup across the shapes your models really use. They should leave behind benchmarks and correctness tests the team can rerun. They also help the team see where time and memory go, so future decisions about hardware and model design rest on evidence.

How you can tell they are good

Ask them to walk through one kernel they optimised, from first profile to final result. A strong candidate names the bottleneck, explains why they chose each change, and shows what failed. They talk comfortably about memory hierarchy, occupancy and precision. Be wary of anyone who quotes a single headline speedup, ignores odd shapes, or cannot say how they checked numerical correctness.

Why it is hard to fill

This is among the scarcest profiles we see. The job titles vary widely, so searching on title misses people. Real skill sits across parallel programming, hardware architecture and numerical methods, and few people have all three at depth. Most good candidates are already inside labs or chip companies, where their work is closely held and their options are many. Public evidence is thin, so judging them takes a specialist's eye.

Where to find them

Look first at engineers from frontier labs, chip companies and accelerator software teams. Strong candidates also come from compiler teams, high performance computing, graphics and game engine work, and systems groups that tune numerical libraries. Titles include performance engineer, CUDA engineer, compiler engineer and runtime engineer. Open source kernel and compiler projects are a good place to see real work, such as merged changes and clear benchmarks.

Indicative pay

Not enough published evidence yet to give a fair range for this role, which is common for work this new. We benchmark it for you as part of a search.

Interview questions for a GPU Kernel Engineer

  1. Tell me about a kernel you made faster. How did you find the bottleneck and what did you change?

    A strong answer starts with profiling data, names the limit (memory, compute or overhead), and links each change to a measured effect. They mention dead ends. Worry if the story is only about final speedup with no diagnosis.

  2. How do you decide whether an operation is memory bound or compute bound, and what do you do about each?

    Look for roofline-style thinking, arithmetic intensity, and practical fixes such as fusion, tiling or better data layout for memory limits. Worry about vague rules of thumb or answers that skip measurement.

4 more questions, with what a strong answer covers, are in the Role Kit.

Job description you can post

We are hiring a GPU Kernel Engineer to make our model operations run faster and cheaper on accelerators. You will work with researchers and systems engineers to profile real workloads, write efficient kernels, and prove they are correct. Your work will directly shape how far our hardware goes.

What you will do

  • Profile model workloads and identify compute, memory and overhead bottlenecks
  • Write and tune kernels in CUDA and Triton for key model operations
  • Build tests that check numerical correctness across representative shapes and precisions
  • Maintain benchmarks that track speedup and catch regressions
  • Work with researchers to shape new operations so they run well on hardware
  • Collaborate with compiler and runtime engineers on how kernels fit the wider system

What you will bring

  • Shown work optimising accelerator kernels, with profiling evidence
  • Strong grasp of parallel programming and memory hierarchy
  • Experience writing CUDA, Triton or comparable kernel code
  • Ability to reason about numerical precision and test for errors
  • Comfort reading compiler output and intermediate representations
  • Clear written explanation of performance trade-offs

Nice to have

  • Contributions to open source kernel or compiler projects
  • Experience with distributed training or inference systems
  • Background in compilers or hardware architecture

Free to use and adapt. If it saves you time, a link back to this page is appreciated.

Skills that matter

  • GPU programmingAdvancedOptimise parallel accelerator kernels
  • Hardware and computer architectureAdvancedDesign computing structures and hardware interfaces
  • Linear algebra and optimisationWorkingReason about matrix operations and objective functions
  • Distributed systemsWorkingDesign reliable work across networked machines

Other roles in aI compilers, kernels and runtime systems

Hiring nowAI Inference Compiler EngineerAn AI Inference Compiler Engineer makes trained models run faster and cheaper on real hardware. They change how model computation is compiled and…Hiring nowML Runtime EngineerAn ML Runtime Engineer builds the scheduling, memory and execution layer that runs models on real hardware. Their work decides whether a model…EmergingAI Performance Portability EngineerAn AI Performance Portability Engineer keeps models fast and accurate across different hardware. They cut dependence on one chip or vendor, which…

Related roles in other families

Hiring nowAI Accelerator ArchitectAI accelerators and hardware co-designAn AI Accelerator Architect decides how a chip or compute system should be built for specific training or inference workloads. Their choices set the…Hiring nowAI Hardware Verification EngineerAI accelerators and hardware co-designAn AI Hardware Verification Engineer proves that accelerator logic and its interfaces behave correctly before silicon is committed. Escaped defects…Hiring nowHardware-Software Co-Design EngineerAI accelerators and hardware co-designA Hardware-Software Co-Design Engineer shapes model software and chip interfaces together, so neither is tuned in isolation. For an AI company, that…EmergingNeuromorphic Computing ResearcherAI accelerators and hardware co-designA neuromorphic computing researcher tests whether event-driven, brain-inspired computation can beat conventional hardware on a real task. The answer…

How we wrote this

Role definitions are our own editorial work, checked against occupational classifications, standards, research and employer pages. Titles vary by company, and this is a description of the work, not a job advert.

← AI compilers, kernels and runtime systemsAll new job families

Start a search

Hiring a GPU Kernel Engineer?

Tell us the work. We find the people already doing it, whatever their job title says, and you meet them already interviewed for your role.