2138 Accelerating DLRM on an OCP M 2 Accelerator

Описание: High Performance and Hardware Acceleration - Recommendation  models  are  widely  deployed  by  a large  number  of  hyperscalars,  taking  up  significant  portions  of the  total  compute  time  in  data  centers.  However,  accelerating these models remains challenging due to their unique architecture with  low  compute  density  sparse  operators.  In  this  work,  we introduce our hardware accelerator for sparse operations, SEAL, which  is  compatible  with  the  OCP  Yosemite  v2  server  platform and  provides  a  performance  improvement  of  up  to 6.2×over  a solution  using  only  CPUs.