GPU Kernel Expert
Listing checked August 29, 2026 · pay as published by Mercor
Overview
This remote contract role puts you inside the evaluation loop for GPU/accelerator kernel tasks generated for a frontier AI lab. You will review assignments built on CUDA, Triton, NKI, and Pallas, checking whether they are numerically sound, correctly scoped, and safe to run. Your written, rubric-based feedback helps decide which tasks are used to train and evaluate the lab's models.
What You'll Do8
- 1Assess kernel tasks for numerical correctness by reviewing tolerance settings and reference-implementation choices.
- 2Run performance checks using profiling tools such as Nsight Compute, ncu, or framework-native profilers to confirm benchmarks are fair.
- 3Verify that each task compiles and executes without hidden runtime issues across target accelerator environments.
- 4Identify common kernel failures, including driver mismatches, out-of-memory states, launch-configuration errors, shape and stride mismatches, and autotuning failures.
- 5Evaluate tasks across multiple categories, including generation from specification, translation between frameworks, hardware migration, debugging, performance optimization, and operator fusion.
- 6Write clear, rubric-based feedback that explains quality scores and points to specific improvements for each kernel task.
- 7Flag any experimental setup that could bias performance comparisons before the task enters a training or evaluation pipeline.
- 8Document findings in a structured written format that a frontier AI lab can use to refine task generation.
Requirements9
- 13+ years of hands-on experience developing, optimizing, or verifying GPU or accelerator kernels in at least two of the following: CUDA, Triton, NKI, or Pallas.
- 2Deep understanding of numerical-correctness standards, including absolute, relative, and ULP tolerances, plus how to select a proper reference implementation.
- 3Proven experience with performance profiling and benchmarking tools like Nsight, ncu, roofline analysis, or framework-native profilers.
- 4Working knowledge of common compilation and runtime problems, such as driver version mismatches, OOM errors, launch configuration failures, shape or stride mismatches, and autotuning inconsistencies.
- 5Practical exposure to at least three kernel task types: generation from specification, translation or lowering, migration across hardware targets, debugging, performance optimization, or operator fusion.
- 6Preferred: experience across both NVIDIA systems (CUDA or Triton) and custom accelerators such as NKI, Pallas, or TPU.
- 7Preferred: background in compiler engineering, MLIR, or intermediate-representation lowering.
- 8Preferred: familiarity with memory-hierarchy optimization, including shared-memory tiling, register pressure, bank conflicts, and coalescing patterns.
- 9Preferred: prior contributions to kernel libraries such as cuBLAS, cuDNN, Triton community kernels, or JAX/XLA custom calls.
Who Should Apply
The ideal candidate is an engineer who enjoys auditing and critiquing kernel implementations more than building them from scratch. If you prefer long hands-on coding sessions over written evaluation and feedback, this role will likely feel like a poor fit. Candidates often get rejected when they cannot show actual profiling results or when they struggle to explain ULP tolerances and reference-implementation reasoning. Another common gap is narrow experience, for example only writing CUDA kernels without touching translation, debugging, or operator fusion.
Salary Insight
The role pays $70 to $90 per hour, which is a strong contract rate for a kernel specialist with 3+ years of experience and multi-framework profici. The final rate depends on your breadth across GPU and custom-accelerator ecosystems and how well you can demonstrate evaluation skills.
Pay and demand for AI Training roles
AggregatedTypical pay
$70/hour
This role
$70–$90/hr
Most AI Training roles pay $42–$98 per hour. This role's pay falls inside that range.
Based on 1,081 similar roles that publish pay · 303 publish only a top rate; those count at the rate they gave
Rates shown per hour. Yearly and monthly pay converted; one-time fees and non-USD pay are not included.
- Live similar roles
- 1,234
- Listed in last 30 days
- 474
- Remote
- 98%
Hiring most right now: micro1 (360) · SME Careers (254) · Mercor (159)
Most requested skills · share of roles
- ai training13%
- llm evaluation11%
- ai evaluation11%
- data annotation9%
Figures from AI Training roles live on NearSkill when this page loaded. A role can close before you apply, so check the listing itself.
Compare your resume against these rolesLocation
Compensation
$70–90/hr
Required Skills
Application Tip
Include a short audit report you have written for a real kernel task, especially one that shows how you identified a benchmark fairness issue or a numerical tolerance flaw.
How NearSkill helps you apply
1. Check your job match
We compare your resume with the job requirements to help you understand how well your experience fits.
2. Continue to the employer’s website
We’ll direct you to the employer’s application portal, where you can complete the application.
3. Get updates by email
If NearSkill is handling follow-up for this role, we’ll email you when there’s an update on your screening.
Want to check your fit first?
See how your resume matches this role before applying.
Check my job matchSimilar open positions
Explore active roles that match your skills and interests.
Mercor
VerifiedTrainium (NKI) Kernel Expert
You will evaluate the quality and correctness of NKI development tasks for a frontier AI lab's model training. Your reviews cover CUDA to NKI migration fidelity, Trainium-specific performance tuning quality, and cross-platform numerical-correctness standards. You will produce rubric-based written feedback that shapes how the lab builds and trains its models on AWS Trainium hardware.
AfterQuery
VerifiedML Inference & GPU Kernel Research Engineer
AfterQuery builds a research cohort of senior ML inference and GPU kernel specialists to test how frontier AI models handle expert-level problems. You will not ship production services. Instead, you design hard scenarios, write reference answers and grading rubrics, then judge model outputs for technical accuracy and domain judgment. The work centers on ML inference, GPU kernel engineering, and large-model deployment, with schedules that stay asynchronous and remote. Candidates with 6 to 10 years in the field and a PhD, or a master's plus a strong research record, who send the requested materials get priority review.
Mercor
VerifiedExpert Interviewer
Mercor supports a frontier AI lab's high-priority technical-expert panel. The interviewer in this role vets shortlisted engineers across GPU kernel development, security & vulnerability research, and ML-compiler / accelerator domains. You run structured 20-minute interviews, score each candidate against a rubric, and provide an evidence-based read. The engagement is part of an AI model training-and-evaluation effort and operates 100% remote.
Talentxo
VerifiedSoftware Engineer (C++ / HPC / GPU Computing)
This role centers on writing high-performance C++ code that runs close to the hardware, with work spanning systems programming, parallel computing, and GPU kernel development using CUDA and HIP. You will design and maintain software for compute-intensive workloads, profile applications to find bottlenecks in compute, memory, and I/O, and deliver measurable speedups. The position involves optimizing memory layout and data movement across the memory hierarchy, plus working with Linux internals such as processes, threads, synchronization, and virtual memory. Engineers who want to understand how code maps to memory, threads, and hardware rather than just using libraries will find the work engaging. The team also evaluates new hardware, toolchains, and optimization techniques together.
AfterQuery
VerifiedSoftware/ML Engineer - LLM Inference
AfterQuery needs senior software and ML engineers to define how well frontier models reason about LLM inference, GPU kernel design, and model-serving architecture. You will craft evaluation scenarios and reference answers for real systems-engineering problems, then assess model outputs against those benchmarks. The contract runs remote at 3 hours per week, so you can schedule work around another job. Work centers on research and evaluation, not shipping production code for AfterQuery.
Meridial
VerifiedTechnical Task Auditor
Independent contractors in this role audit tasks that train and evaluate AI systems. You pick one specialty from eight areas, such as GPU kernels, AWS Trainium, NKI, CVE and application security, machine learning, AI-assisted developer workflows, AWS Serverless and Infrastructure as Code, Kubernetes, or SWE-Bench. Your job is to judge whether each task is accurate, realistic, solvable, and reproducible, and whether reliable tests and evaluation criteria support it. When something falls short, you write feedback that other experts can act on. Not every specialty is required, and you complete a technical assessment for your chosen area during selection.