Skip to content

Latest commit

 

History

History
51 lines (39 loc) · 2.39 KB

PM3 Power Modeling and Power Management for Processing-in-Memory-HPCA'18-LX.md

File metadata and controls

51 lines (39 loc) · 2.39 KB

Paper title:

PM3: Power Modeling and Power Management for Processing-in-Memory

Publication:

HPCA’18

Problem to solve:

When the Processing-in-Memory (PIM) is prevalent in accelerating data-intensive application, there are two problem becoming more and more severe in PIM application. Firstly, the power consumption at the ideal peak throughput in previous PIM designs may exceed its power supply beyond the DRAM thermal tolerance capability with passive heat sink, which may have potential power supply failure and memory reliability problems. Secondly, the power supply is not fully utilized in this case. Therefore, it is important to have a comprehensive quantitative study of the power modeling and power management for such PIM architectures.

Major contribution:

  1. This paper models the relationship between bandwidth and power consumption, to facilitate early stage PIM power system designs, and then proposed the Power-Aware Subtask Throttling(PAST) technique to reduce PIM power requirement.

  2. This paper proposes the Processing-Unit Boost (PUB) technique based on a greedy algorithm, to improve the performance of processing units, and the Power Sprinting (PS) technique to improve the energy efficiency when the power supply is dynamically adapted.

  3. This paper proposes evaluating both HMC-based and RRAM-based PIM designs for their targeted applications. Also, via evaluations it demonstrated that NVMs and 3D DRAM have their own advantages in different bandwidth usage cases to be selected as PIM memories.

Lessons learnt:

  1. Dynamic power management may provide an insight into improving the energy efficiency and performance. Ideal method to consume the power is to put every drop of power on the crucial point and break the power wall. However, this is hard to target. So checking out the main bottleneck in a system comes with gains, and then design a scheme to address problem depending on this bottleneck model. This may be based on that domain-specific power consuming law, that is, there is a similar distribution of power usage within a field. That may be an interesting direction.

  2. To address the challenges in power management, let the system be aware of the power consuming throttling is an effective way to provide the points to optimize the performance.