Optimizing a CUDA Matrix Multiply Step by Step, From Naive to Tensor Cores
How a CUDA SGEMM goes from 1% to over 90% of cuBLAS: coalescing, shared memory tiling, register blocking, double buffering, warp tiling and tensor cores.
Deep dives, practical guides, and incident analyses from engineers who build Safeguard. No fluff, no vendor FUD — just what you need to ship secure software.
How a CUDA SGEMM goes from 1% to over 90% of cuBLAS: coalescing, shared memory tiling, register blocking, double buffering, warp tiling and tensor cores.
An NVIDIA SM is four partitions, each with a warp scheduler, a 64 KB register file and a tensor core. How they hide latency, and how to compute occupancy.
CUDA runs a kernel as a grid of thread blocks, split into 32-thread warps on SMs. Here is how SIMT, memory spaces, coalescing and sync really work.
Photolithography prints chip patterns with light through a mask into photoresist. How contact printing became 193 nm immersion, multi-patterning and overlay.
EUV scanners make 13.5 nm light from tin plasma, bounce it off Mo/Si mirrors and a reflective mask, and scan wafers to sub-nanometre precision.
A GPU turns triangles into pixels with thousands of parallel shader threads. Follow a draw call through rasterization, ROPs, ray tracing and compute.
Vedic, Trachtenberg, soroban, Chisanbop, peasant, lattice, base 60, base 20 and champion methods run on the same 7 problems, with step and memory counts.
Line multiplication draws each digit as sticks and counts the crossings per zone. Worked on four problems, proved with algebra, and its name checked.
How the Trachtenberg system multiplies by 11, 12, 5 to 9 and any number, adds and divides, worked step by step on fixed problems, with its real history.
Weekly insights on software supply chain security, delivered to your inbox.
Your first fix PR is minutes away.
No sales call required, even your agent can complete the purchase over MCP.