• core-underutilization | RTX 4090, TCU | | | GE-SpMM | 2020 | | | RTX 2080 (DRAM 484GB/s) | |

Background

CSR

SpMV in CSR

SpMM in CSR

Paper List

DTC-SpMM

GE-SpMM

Acc-SpMM

core concept

BW and Latency

• BW vs latency, not as conflicting concepts

1) : stalled by waiting data which requires long access latency (have opportunity for prefetching)
2) : stalled by waiting data which requires long access latency, saturating all the DRAM BW (have to increase the arithemetic intensity)
3) : stalled by waiting data which saturates all the DRAM or cache BW
4) : no stall

• solutions
  • for BW bounded problem : compression, changing algorithm to re-use more (increasing the arithemetic intensity),
  • for lantecy bounded problem : prefetch,