Processing-in-memory in High Bandwidth Memory (PIM-HBM) Architecture with Energy-efficient and Low Latency Channels for High Bandwidth System

In this paper, for the first time, we propose a processing-in-memory in high bandwidth memory (PIM-HBM) architecture for high bandwidth systems with low dynamic random-access memory (DRAM) access costs. The main concept of the proposed PIM-HBM architecture is to embed processing units into a logic base of high bandwidth memory (HBM) to decrease the energy consumption and latency of interconnections as the physical length between core and DRAM decreases. To verify the proposed PIM-HBM architecture, we designed on-chip and on-interposer I/O channels using a CMOS 0.18 µm process. We extracted channel parasitic using an electromagnetic (EM) solver and performed a SPICE simulation to compare the system performance of the proposed architecture with the conventional HBM. As a result, the performance of the proposed PIM-HBM architecture is successfully verified by reducing energy consumption and latency of interconnections by 77 % and 79 % compared to the conventional HBM system.

[1]  Joungho Kim,et al.  Signal Integrity Design and Analysis of Silicon Interposer for GPU-Memory Channels in High-Bandwidth Memory Interface , 2018, IEEE Transactions on Components, Packaging and Manufacturing Technology.

[2]  Joel Emer,et al.  Eyeriss: a spatial architecture for energy-efficient dataflow for convolutional neural networks , 2016, CARN.

[3]  Reum Oh,et al.  18.2 A 1.2V 20nm 307GB/s HBM DRAM with at-speed wafer-level I/O test scheme and adaptive refresh considering temperature distribution , 2016, 2016 IEEE International Solid-State Circuits Conference (ISSCC).

[4]  Xi Chen,et al.  Ground-referenced signaling for intra-chip and short-reach chip-to-chip interconnects , 2018, 2018 IEEE Custom Integrated Circuits Conference (CICC).

[5]  Mahmut T. Kandemir,et al.  Scheduling techniques for GPU architectures with processing-in-memory capabilities , 2016, 2016 International Conference on Parallel Architecture and Compilation Techniques (PACT).