Robert Zieliński / Simulators / Computer Science / Hardware / GPU

GPU SIMT execution — warps · latency hiding · divergence · coalescing · RTX 4090

occ — warps — util —

Speed cycle 0

GPU

SM · idle cycle 0

Pick a scenario or open the kernel editor to launch a kernel.

Kernel editor

A Python-subset kernel — every one of the 32 lanes in a warp runs it for real. Use tid (global thread index) and the input arrays by name. Supported: if/elif/else, for i in range(), while, def, lists & dicts, arithmetic, and CUDA intrinsics (cuda.grid(), cuda.shared.array(), cuda.syncthreads(), cuda.atomic.add()).

Kernel source

Input arrays (name = [..] or name = zeros(N)) Launch geometry

blocks × threads

Registers / thread (drives occupancy)

SM latency hiding · Lanes divergence · Memory coalescing · Hierarchy grid/block/thread · CPU vs GPU roofline issue active stalled divergent shared mem