SahibnoorSingh Chahal
I keep going one layer deeper. Dashboards led me to the pipelines underneath them, data led me to machine learning, and now I'm writing GPU kernels to see what makes models fast.
Ongoing since Apr 2026
Learning GPU performance by writing kernels
Testing whether hardware profiling signals from Nsight Compute — occupancy, memory throughput, warp stall reasons — can guide kernel tuning in Triton. Softmax, tiled matmul and fused bias+GELU+dropout kernels, checked against PyTorch eager and torch.compile on an RTX 3060, each classified as memory- or compute-bound.
3.3×fused bias + GELU + dropout vs. PyTorch eager