We're seeking a GPU Performance Engineer to squeeze every last FLOP from our H100 infrastructure and optimize our model serving stack to its absolute limits.
The RoleYou'll be our performance optimization expert, using advanced profiling tools to identify bottlenecks and implementing solutions that achieve 5-10x speedups. From writing custom CUDA kernels to eliminating cold start latency, you'll ensure our infrastructure delivers world-class performance. This role is perfect for someone who gets excited about microsecond optimizations and pushing hardware to its theoretical limits.
Key Responsibilities- Profile and optimize GPU workloads using Nsight Systems, nvprof, and custom instrumentation
- Write high-performance CUDA and Triton kernels for critical model operations
- Optimize cold start latency from seconds to milliseconds for our serving infrastructure
- Tune memory access patterns, kernel fusion, and GPU utilization
- Collaborate with ML engineers to optimize model implementations
- Debug performance issues across the full stack from application to hardware
- Implement custom memory pooling and allocation strategies
- Share optimization techniques and build performance culture across teams
Qualifications- Bachelor's or Master's degree in Computer Science, Electrical Engineering, or related field
- 5+ years systems programming experience with 3+ years focused on GPU optimization
- Expert proficiency with GPU profiling tools (Nsight Systems, nvprof)
- Strong CUDA programming skills with production kernel development
- Deep understanding of GPU architecture (memory hierarchy, SMs, warps)
- Track record of achieving significant performance improvements (5-10x)
- Experience with Python and C++ in production environments
We Value- Experience with Triton kernel development
- Knowledge of CUTLASS or similar high-performance libraries
- Background in ML-specific optimizations (attention, transformers)
- RDMA/InfiniBand optimization experience
- Contributions to GPU libraries or frameworks
- Low-level debugging skills (PTX/SASS reading)
