Skip to content

CUDA

Compute Unified Device Architecture is a parallel computing platform and application programming interfac

1. Graphics Processing Unit

  • SIMD
  • low frequency (1.4 GHz), but many cores (8k+)
  • powerful: 29.7 TFlops
  • high speed of DRAM: 760 GB/s (total)
  • mem. hierarchy:
    • regester
    • shared mem. : PBSM: per-block shared mem.
    • global mem.
    • const mem. ?

GPU

  • Streaming Multiprocessor
    • Warp: locked step for all cores
      • Core

Host: CPU + CPU Mem.

Device: GPU + GPU Mem.

Kernel function

kernel = grid -> thread blocks -> threads

  • thread blocks should be independent; they are parallel
  • thread blocks are mapped to SMs

grid? how it combined with hardware

  • limited number of threads in each thread block: 1024; use multiple blocks

  • limited total number of threads

  • __device__ can only be called on GPU

  • __global__ called on host and runs on GPU

Compilation Process (by nvcc)

  1. seperate
  2. generic: generate PTX
  3. specialized

coalesced mem. access

Last update: April 19, 2022
Authors: Co1lin