Several optimization methods of half-precision general matrix vector multiplication (HGEMV) using CUDA core.

An implementation of SGEMV with performance comparable to cuBLAS.

PCCX is an open NPU architecture for memory-bound Transformer inference on edge FPGAs, focused on GEMM/GEMV, KV-cache, W4A8 quantization, and custom ISA scheduling.

Highly optimized DGEMV on CPU with both serial and parallel performance better than MKL and OpenBLAS.

Matilda is a library to repeatedly multiply a constant matrix with a variable vector

Measure and visualize why LLM inference is slow: bottleneck analysis, model dissection, KV-cache, GEMM/GEMV, quantization, and memory-bound decoding.

CUDA/C++ matrix-vector and matrix-matrix kernels with naive, shared-memory tiled, and warp-coalesced variants, plus GFLOPS benchmarking vs CPU.

🧮 CereMath is a library of Machine Learning kernels for the Wafer Scale Engine (WSE) from Cerebras. Made as a thesis project