✳flâneur — a map of the web's best reading

BrrrViz

brrrviz.com · 1,419 words · saved by 1 readers

A Graphics Processing Unit (GPU) is thousands of highly-parallel, small processors organized into synchronized groups, all sharing a strict memory hierarchy. Click either die to compare compute, control, and cache. CPUs spend area on cache and control for a few wide, sequential cores. GPUs spend it on ALU density and lean control for massive parallelism. 1 core: only one can help 32 ALUs, but only 1 can run at a time When each element depends on the previous one, parallelism can't help. The GPU is stuck using a single ALU, and a single GPU ALU is slower than a CPU core. All those extra ALUs sit disabled because the dependency chain forbids them from running. SELECTED Integer pipelines for 32-bit ops: addresses, offsets, predicates, and loop counters. They sit next to the FP CUDA cores in the same block but are not the same thing as FP32/FP64 CUDA cores. Each tile is one streaming multiprocessor (SM). L2 and VRAM below are shared by the whole chip. Click any SM to open the SM diagram. C

A Graphics Processing Unit (GPU) is thousands of highly-parallel, small processors organized into synchronized groups, all sharing a strict memory hierarchy. GPUs don't run one thread at a time, they launch thousands of parallel processes. When threads in a warp take different paths, the GPU serializes execution. Learn the causes and mitigations. Threads in a block must coordinate access to shared memory. Learn why races happen, how barriers order access, and what they cost. Shared memory is striped across 32 banks. When multiple threads in a warp hit the same bank with different addresses, ac

Explore this link on the map →

saved by

related reading