Supporting Massive DLRM Inference Through Software Defined Memory

10/21/2021

∙

by Ehsan K. Ardestani, et al.

∙

Facebook

∙

Deep Learning Recommendation Models (DLRM) are widespread, account for a considerable data center footprint, and grow by more than 1.5x per year. With model size soon to be in terabytes range, leveraging Storage ClassMemory (SCM) for inference enables lower power consumption and cost. This paper evaluates the major challenges in extending the memory hierarchy to SCM for DLRM, and presents different techniques to improve performance through a Software Defined Memory. We show how underlying technologies such as Nand Flash and 3DXP differentiate, and relate to real world scenarios, enabling from 5 power savings.

READ FULL TEXT

Supporting Massive DLRM Inference Through Software Defined Memory

Data Center Interconnects at 400G and Beyond

An Analysis of Deep Neural Network Models for Practical Applications

Characterizing the Power Cost of Virtualization Environments

Survey of Machine Learning Accelerators

Efficient Synthesis of Compact Deep Neural Networks

Improved Ahead-of-Time Compilation of Stack-Based JVM Bytecode on Resource-Constrained Devices

SmartWatts: Self-Calibrating Software-Defined Power Meter for Containers

Supporting Massive DLRM Inference Through Software Defined Memory

Related Research

Data Center Interconnects at 400G and Beyond

An Analysis of Deep Neural Network Models for Practical Applications

Characterizing the Power Cost of Virtualization Environments

Survey of Machine Learning Accelerators

Efficient Synthesis of Compact Deep Neural Networks

Improved Ahead-of-Time Compilation of Stack-Based JVM Bytecode on Resource-Constrained Devices

SmartWatts: Self-Calibrating Software-Defined Power Meter for Containers