A new technical paper titled “Achieving the Performance of All-Bank In-DRAM PIM With Standard Memory Interface: Memory-Computation Decoupling” was published by researchers at Korea University.
“This paper proposed the memory-computation decoupled PIM architecture to provide the performance comparable to the all-bank PIM while preserving the standard DRAM interface, i.e., DRAM commands, power budget, timing constraints, etc. For achieving our goal, we introduced two PIM execution phases: memory and computation. The memory phase follows the per-bank PIM execution method. The computation phase broadcasts the bank-shared operands to all banks, thus making their PIM engines perform the computation simultaneously without accessing the bank data arrays,” states the paper.
Find the technical paper here. Published August 2022.
Y. Paik, C. H. Kim, W. J. Lee and S. W. Kim, “Achieving the Performance of All-Bank In-DRAM PIM With Standard Memory Interface: Memory-Computation Decoupling,” in IEEE Access, vol. 10, pp. 93256-93272, 2022, doi: 10.1109/ACCESS.2022.3203051. Creative commons license here.
Leave a Reply