Accelerator Kernel and Shader Reviewer
Job Description
IXO is engaging specialists to evaluate GPU implementations whose speed and numerical behavior can both be demonstrated. Your contribution is technical work: make the reasoning inspectable, identify substantive errors and provide evidence that supports a reliable assessment.
Work you will do
• Develop, inspect or optimize kernels and shaders in the accelerator ecosystem relevant to the assignment.
• Use profiling evidence to locate throughput, memory, launch or other performance limits, and explain the consequences of an implementation change.
• Check host-side C+• integration and numerical results across the required hardware or precision settings; document reproducible benchmark findings.
Required background and routes
• Kernel-audit route: at least three years implementing, tuning or verifying accelerator kernels, with experience in two or more of Pallas/JAX, NKI, Triton and CUDA.
• Assess absolute/relative/ULP tolerances against a suitable reference; profile performance with tools such as ncu, Nsight or framework profilers and reason about roofline limits.
• Recognize driver, launch, allocation, shape/stride and autotuning failures. Bring experience in at least three task families: specification-to-code, framework lowering, hardware migration, debugging, optimization or operator fusion.
Preferred background
• Cross-platform accelerator work, MLIR/compiler expertise, memory tiling/register/bank/coalescing optimization, or contributions to cuDNN, cuBLAS, Triton kernels or JAX/XLA calls.
• The CUDA/shader route favors strong C++, CUDA performance work, GLSL and WebGPU, GPU-profiler use and clear hardware-generation comparisons.
Deliverables
Submit the completed technical artifact or assessment with its supporting evidence, explicit assumptions, reproducible checks where applicable, and concise reasons for each material judgment. Address review findings within the agreed scope.
Location and schedule
A regional track is open to residents of United States. Remote assignments are scheduled by agreement, with no guaranteed weekly volume. Availability planning can include 40 hours per week depending on the track. IXO confirms the applicable timing before work.
Pay and working terms
$75 • $105/hr USD. The agreed hourly rate, scope, schedule and acceptance criteria are confirmed before work starts. Applying does not guarantee an assignment. Use public, licensed or otherwise authorized material only; do not submit confidential employer information, personal data or restricted research.