Karnbir Khera

I enjoy profiling GPU kernels from first principles, tracing performance behavior back to hardware architecture using Nsight Compute.What started as a vector addition project became a month-long investigation into L2 cache write-validate policies, designing micro-benchmarks to isolate cache behavior, tracing root cause to a 2018 arXiv paper, and validating the findings across Ada Lovelace (RTX 4060) and Blackwell (B200) architectures. That work became a 40-page technical report documenting roofline analysis, warp stall behavior, occupancy tradeoffs, and vectorization width decisions across six kernel variants.Currently competing in the NVIDIA MLSys 2026 FlashInfer Sparse Attention Competition (Track B), building toward sparse attention kernel optimization on B200 hardware in pure CUDA, prioritizing hardware intuition over abstraction at every step.Building toward GPU kernel engineering for deep learning optimization.