용어 및 간단한 개념 정리

MSHR : (Miss-status Holding Registers)

Associative comparison logic merge simultaneous requests for the same cache block, the number of outstanding misses  that can be supported is typically small.

[출처 : analyzing cuda workloads using a detailed gpu simulator, ispass 2009]

CTA : Cooperative Thread Array

 

CUDA 프로그래밍 모델에서의 GPU 메모리 영역의 종류

-> global, local, constant, texture, shared

 

개수 ( Architecture : GF100 , Compute capability : version 2.x )

1warp 당 thread 32개

1block 당 thread 1024개

1block 당 warp 32개

1SM 당 CUDA core 32개

Maximum # of threads per block :  1024개

 

기본 구조