CUDA Engineering Expert

$300/taskrate as listed

Mercor is contracting GPU kernel specialists to review, optimize, and reason about CUDA/HIP kernels for a leading AI lab's training data and model evaluation work. You'll work from profiler traces — occupancy, L2 hit rate, memory throughput — and document why a given optimization does or doesn't help, at an observed $80–100/hr for roughly 20+ hours per week.

Summary, eligibility guidance and screening preparation are prepared by Sidequest from the platform's posting. Check the source listing before applying.

What the work looks like

Most tasks start with a kernel and a profiler report rather than a spec. You'll read the implementation, form a hypothesis about the bottleneck — memory-bound vs. compute-bound, occupancy limits, bank conflicts, poor coalescing, launch overhead — and either produce an optimized version or evaluate a model-generated one. A large share of the work is judgment on AI output: deciding whether a proposed rewrite is actually faster, whether it's numerically equivalent, and whether the model's stated reasoning about a metric is correct. Written justification matters as much as the code; the lab wants to know when L2 hit rate is a red herring, not just when it's high.

You will not always get deep algorithmic context. Part of what's being screened for is the ability to optimize a kernel you didn't write, for an algorithm you haven't studied, using the hardware behavior alone as your guide.

What the screen measures

Fluency, not familiarity. Expect follow-ups that go past textbook answers: what limits occupancy on a specific SM configuration, when shared memory staging costs more than it saves, how tensor core tiling constrains your data layout.
Profiler literacy. NSight Compute experience is a plus on paper and close to essential in practice. Be ready to explain which counters you actually trust and why.
C++17 competence. Templates, constexpr, move semantics — the code you touch is modern C++, not C with kernels.
Honest scoping. Claiming Blackwell, inline PTX, or HIP depth you don't have tends to surface immediately under follow-up questioning.

Logistics

Fully remote, asynchronous, contract-based through Mercor. The stated minimum is 20 hours per week; contributors report work arriving in batches with flexible scheduling inside that commitment. Pay in the $80–100/hr range has been observed for this listing — rates vary by demonstrated depth, assessment performance, and project phase, and are not guaranteed. Applications begin with a resume review and may include a short technical assessment before onboarding.

What they look for

Fluency in at least one GPU programming model (CUDA, HIP, HLSL, GLSL, Slang) with 1+ year professional or graduate research experience
Core C++ through C++17, including templates and modern idioms
Working command of GPU profiler metrics — occupancy, L2 hit rate and throughput, memory coalescing — and how to act on them
Ability to identify and fix kernel bottlenecks without deep prior context on the underlying algorithm
Python and Git for tooling, benchmarking, and submission workflow
Plus: NSight Compute, inline PTX, tensor core optimization, or NVIDIA Blackwell tuning experience

Before you apply

Would you pass this screen?

Drawn from what this listing actually requires. Answered in your browser — nothing is sent to us, and we do not store your answers.

Have you written and executed GPU kernels in CUDA, HIP, or another GPU kernel language (Slang, HLSL, GLSL, or similar)?
CUDA or GPU programming language · required
Do you have working fluency in C++17, including features like templates, auto, lambdas, and modern standard library usage?
C++ through C++17 · required
Do you have at least 1 year of professional or graduate-level research experience writing or optimizing GPU code?
1 year GPU experience · required
Are you familiar with GPU profiler metrics such as occupancy, memory throughput, cache hit rates, and latency, and have you used them to guide optimization decisions?
Profiler metrics knowledge · required
Are you available to work at least 20 hours per week on this contract?
20 hours per week availability · required
Do you have working knowledge of Python and Git version control?
Python and Git · preferred

0 of 6 answered

Included with this listing

Interview prep — 11 questions

Matched to what this platform's screening actually measures. Answer each out loud in 60–90 seconds: claim → concrete example → implication.

What it tests: Whether they can ground profiler output (L2 cache, occupancy, throughput) in concrete optimization decisions, and whether they're comfortable explaining GPU behavior without deep algorithmic context.
A strong answer: A specific kernel with a named bottleneck identified by profiler metrics (e.g., L2 cache misses, low occupancy, memory bandwidth saturation), the change made to address it, and the measured improvement. Profiler output should drive the narrative.
Where people lose it: Vague language like 'made it faster' or 'improved cache' without naming the profiler signal that revealed the problem, or claiming they optimized a kernel but can't articulate which metric guided the work.