GPU Kernel Performance Engineer
GPU Kernel Performance EngineerAI workload efficiency | Mountain View / South Bay | OnsiteThe opportunity Acceler8 Talent is partnering with an early-stage AI software startup building technology to improve the performance and energy efficiency of AI workloads. The team is hiring a GPU Kernel Performance Engineer who enjoys understanding why a workload is slow and changing the code that determines how it runs. You will move between profiling, GPU architecture and implementation, connecting low-level changes to measurable workload improvements. There is room to optimize existing kernels, develop new approaches and investigate the surrounding runtime or execution system when that is where the bottleneck sits. Technical scope The work begins with a specific workload and a clear account of where execution time or resources are being consumed. You will examine how threads are organized, how data moves through the memory hierarchy and how synchronization and launch behavior affect the larger application. Optimizing an existing kernel is valuable when you can explain the mechanism, validate correctness and show that the gain matters in the surrounding system.
What you will do:
- Profile GPU workloads and isolate memory, compute, synchronization and launch overhead bottlenecks.
- Implement or modify GPU kernels and validate both numerical correctness and performance.
- Improve memory access patterns, parallel execution and hardware utilization through architecture-aware changes.
- Investigate techniques such as fusion, better data movement and occupancy tuning where profiling supports them.
- Measure the effect of changes on representative workloads and explain the trade-offs.
- Collaborate across compiler, runtime and systems work to improve end-to-end execution efficiency.
- Design repeatable profiling and benchmarking experiments that separate kernel execution cost from launch, synchronization and surrounding runtime overhead.
- Analyze memory access patterns, coalescing, bandwidth use, occupancy and register pressure to identify the constraint that actually limits performance.
- Evaluate changes to work partitioning, data reuse and fusion, balancing reduced memory movement against execution and resource constraints.
- Compare optimized implementations against an appropriate reference across representative shapes or inputs, checking numerical behavior as well as speed.
- Document the conditions under which an optimization helps and investigate where the bottleneck moves after the change.
- What you will bring
- Substantial personal contribution to GPU kernel implementation or optimization.
- Strong understanding of GPU execution, memory hierarchy and performance constraints.
- Ability to connect profiling evidence to specific code changes, measured gains and correctness checks.
- Motivation to investigate difficult performance problems and work beyond a single layer of the stack.
- Ability to reason from GPU architecture and profiling evidence to a specific implementation change.
- Practical experience measuring improvements consistently and identifying misleading or incomplete benchmark results.
- Comfort investigating interactions with compilers, runtimes or application execution when kernel-level changes alone do not explain the result.
- Useful experienceCUDA development and GPU profiling tools, including experience interpreting memory and execution metrics.
- Compiler, runtime, HPC or systems performance work that includes meaningful GPU kernel implementation.
- Experience evaluating workload performance or energy-efficiency trade-offs across more than one layer of execution.
- Substantial technical projects can demonstrate fit for early-career engineers; experienced candidates should show independently owned optimization work.