Scribd
Upload
28 views

SIMDRAM A Framework For Bit-Serial SIMD Processing Using DRAM

This document summarizes a paper that proposes SIMDRAM, a framework for bit-serial SIMD processing using DRAM. SIMDRAM aims to support complex arithmetic operations to enable more widespread adoption of in-memory computing. It expands on prior work by implementing operations using only majority and NOT logic. This reduces memory access bottlenecks. However, using dedicated DRAM rows for computation could increase row hammer attacks. The paper strengths are addressing memory latency issues but weaknesses are not replacing conventional systems entirely and controller overhead. Ideas for improvement include using alternative emerging memories instead of DRAM and exploring transposition unit alternatives.

Uploaded by

Abhay Shriram
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
28 views

SIMDRAM A Framework For Bit-Serial SIMD Processing Using DRAM

This document summarizes a paper that proposes SIMDRAM, a framework for bit-serial SIMD processing using DRAM. SIMDRAM aims to support complex arithmetic operations to enable more widespread adoption of in-memory computing. It expands on prior work by implementing operations using only majority and NOT logic. This reduces memory access bottlenecks. However, using dedicated DRAM rows for computation could increase row hammer attacks. The paper strengths are addressing memory latency issues but weaknesses are not replacing conventional systems entirely and controller overhead. Ideas for improvement include using alternative emerging memories instead of DRAM and exploring transposition unit alternatives.

Uploaded by

Abhay Shriram
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

A review on ‘SIMDRAM: A Framework for Bit-Serial SIMD Processing

using DRAM [1]’


-Abhay Shriram
2019AAPS0282H
-Divakarla Vamsi Krishna
2019A8PS0651H

1) Motivation behind the work-Modern applications use more and more data(eg.Neural Networks),
which increases the energy and delay costs of computing in conventional computer architectures. More
than 60% of the energy used by current systems is used to move enormous volumes of data via
bandwidth-limited memory channels from memory to the CPU. The processing-in-memory (PIM)
paradigm reduces (and in some cases eliminates) the requirement to transfer data between memory and
the processor in order to minimise these costs.

2) Problems addressed in the paper -A small number of fundamental operations have been proposed
for processing-using-DRAM (i.e., logic operations, addition). However, support for more sophisticated
operations is required to enable the complete adoption of processing-using-DRAM.The latency due to the
memory bottleneck is reduced o a large extent.This paper is an extension of the AMBIT methodology
which extends to support complex arithmetic operations.

3) Goal of the paper and key ideas presented -This paper proposes SIMDRAM, a general-purpose
processing framework using DRAM that (1) permits the effective implementation of complex operations
and (2) offers a flexible method to support the implementation of any user-defined operation.Prior works
(AMBIT) implement fundamental logical operations via majority operations. The objective in-DRAM
calculation is then implemented using these fundamental operations as its fundamental building pieces.
By only employing the logically complete set of majority (MAJ) and NOT operations to conduct
in-DRAM calculation, SIMDRAM expands the application of the majority operation.

4) Strengths and weakness of the paper-THE MAJOR MEMORY- BOTTLENECK latency issues
have be addressed in the above paper,but the paper is suggested not to completely replace the
conventional system,it is proposed to use this technology with the existing technology.With the above
hybrid system SIMDRAM uses a vertical layout and the conventional memory layout uses a horizontal
memory layout.A controller is used to address this problem,but causes an overhead in doing
so.RowHammer attacks [2] may become more frequent as a result of SIMDRAM and other similar
in-DRAM computing technologies that use dedicated DRAM rows to execute calculation.

5) Ideas of Improvement- Use of other emerging memory methodologies instead of DRAM to


implement the SIMD architecture. Exploring alternatives for the transposition unit
References
[1] Hajinazar, N., Oliveira, G. F., Gregorio, S., Ferreira, J. D., Ghiasi, N. M., Patel, M., ... & Mutlu, O.
(2021, April). SIMDRAM: a framework for bit-serial SIMD processing using DRAM. In Proceedings of
the 26th ACM International Conference on Architectural Support for Programming Languages and
Operating Systems (pp. 329-345).
[2] P. Frigo, E. Vannacci, H. Hassan, V. van der Veen, O. Mutlu, C. Giuffrida, H. Bos,
and K. Razavi, “TRRespass: Exploiting the Many Sides of Target Row Refresh,”
in IEEE S&P, 2020.

You might also like