20 Nov
|
Spydra
|
Hyderabad
Positions: 1 |
Experience:
5–9 years
About the role
You will optimise GPU compute and memory operations to lower inference latency and cost per token.
Key responsibilities
• Develop and optimise CUDA C++, Triton and, where relevant, ROCm/HIP kernels.
• Optimise attention, matrix multiplication, normalisation and inference operators.
• Improve memory coalescing, register usage, cache behaviour, occupancy and kernel launch efficiency.
• Profile bottlenecks with Nsight Compute/Systems or equivalent tools.
• Evaluate FlashAttention, PagedAttention, fused kernels and low-precision FP8/FP4/INT4 inference where supported.
• Validate numerical correctness and benchmark against established implementations. Required technical skills CUDA C++, Triton, PyTorch, GPU architecture and memory hierarchy, Nsight Compute, Nsight Systems. ROCm/HIP strongly preferred for at least one position. Expected deliverables Validated optimised kernels, reproducible speedups and improved GPU efficiency.
📌 Senior Gpu Kernel Optimization Engineer Hyderabad
🏢 Spydra
📍 Hyderabad