Deep Learning on GPUs: Successes and Promises

By Sparsh Mittal

August 27, 2019

The rise of deep-learning (DL) has been fueled by the improvements in accelerators. Accelerators allow DL models to crunch a large amount of data, which is vital for them to achieve high accuracy. In fact, AlexNet, the famous winner of the ILSVRC 2012 competition, was trained on GPUs. GPU continues to remain the most widely used accelerator for DL applications, due to several of its features, such as high performance, continued improvements in its architecture and software-stack, ease of programming using high-level languages such as CUDA and availability of GPUs in cloud.

“Accelerating DL models” is chasing a moving target

As DL models are becoming more pervasive and accurate, their compute and memory requirements are growing tremendously. For example, training a deep neural network (DNN) takes a large amount of time, e.g., 100-epoch training of ResNet-50 on ImageNet dataset on one M40 GPU requires 14 days. Similarly, during inference, meeting the latency targets while achieving high data-reuse and throughput is a major challenge.

Extracting last bit of performance from GPU

While treating GPU as a black box is a convenient abstraction of DL researchers, even simple architectural optimizations can boost the performance of GPU significantly. For example, since the input-data to DNN remains unchanged, it can be stored in the constant cache. The weights can be loaded in shared memory to avoid incurring the penalty of accessing global memory. Also, partial sums can be stored in the register file to achieve efficient accumulation.

In fact, architecture-oblivious techniques run the risk of losing their theoretical benefits. For example, although weight pruning is expected to increase performance by virtue of reducing the model size of a DNN, on GPUs, pruning actually harms the performance of DNNs. This is because weight pruning makes the DNN sparse, which requires sparse matrix-multiplication (MM). However, optimizations such as memory-coalescing and matrix tiling cannot be performed on sparse MM. To address this inefficiency, researchers suggest doing “node pruning,” and not “weight pruning” on GPU.

Node pruning does not make the network sparse, and although it brings a smaller reduction in model size than weight pruning, it achieves higher throughput by more effectively utilizing the massive resources of GPUs.

Similarly, optimizing data-layouts, batching, and data-reuse is important to get high performance. Also, since convolution can be performed in multiple ways such as FFT, Winograd, lowering (matrix-multiplication) or direct convolution, the choice of the right strategy is essential. The recent survey paper I’ve written with Shraiysh Vaishay reviews many techniques for optimizing DL on GPUs.

Utilizing both CPU memory and GPU memory

DNN training requires a significant amount of memory, which may exceed the memory capacity of a single GPU. For example, training VGG-16 with a batch size of 256 requires 28GB memory, which is larger than the 12GB memory capacity of Titan X.

To alleviate the memory bottleneck issue, the memory resources of CPUs can be used. In the back-propagation algorithm, the feature maps of a layer, which are produced during the forward-propagation phase, are later reused during the backward-propagation phase of the same layer. Since current machine-learning frameworks allocate the memory for accommodating the needs of all the layers, these feature maps stay in GPU memory for a long time without getting used. To alleviate this inefficiency, feature maps not required by the current layer in the forward-propagation phase are offloaded to CPU memory and released from GPU memory. During the backward propagation phase, these feature maps are fetched from CPU memory to GPU memory just before the processing of that layer. Evidently, the GPU memory management techniques and high-bandwidth interconnect such as NVLink can play a significant role in accelerating training of DNN workloads.

HPC is vital for AI

Distributed computing over a cluster of GPUs can reduce the training time of DNNs significantly. For example, researchers from SenseTime Research and Nanyang Technological University, Singapore have trained AlexNet over ImageNet dataset in just 1.5 minutes. They have used a cluster of 64 machines, each with 8 Volta GPUs. They also perform a range of optimizations at all levels of abstraction, such as using NVIDIA’s NCCL communication library and storing parameters and gradients in half-precision (FP16). Also, they overlap the communication of gradient of one layer with backward propagation of subsequent layers, combine multiple allreduce operations into one operation to reduce the memory copy overhead and intelligently transmit only those gradients that exceed a threshold.

Similarly, researchers from Sony corporation have trained ResNet-50 in just 2 minutes using 3,456 Volta GPUs. This “race to train DNNs” is no less exciting than the “race to the moon” seen in the 1960s! On a more serious note, the DNN training performance can be a more meaningful metric for HPC systems than the peak performance metrics such as Exaflop. This has already led to the creation of benchmarks such as DawnBench and MLPerf.

AI accelerator future promises to be exciting

While the general-purpose nature of GPU makes it useful for a broad range of applications, it also precludes thorough optimization of GPU architecture for AI applications. In this regard, custom-made AI accelerators such as Google’s tensor processing unit (TPU) are in a vantage position. It remains to be seen whether the future trajectory of GPU architecture will see revolutionary or evolutionary changes. It will be also interesting to see how well the next-generation GPU strikes a balance between the conflicting goals of special-purpose and general-purpose computing, and how well it competes with the other AI accelerators.

About the Author

Sparsh Mittal received the B.Tech. degree in electronics and communications engineering from IIT, Roorkee, India and the Ph.D. degree in computer engineering from Iowa State University (ISU), USA. He worked as a Post-Doctoral Research Associate at Oak Ridge National Lab (ORNL), USA for 3 years. He is currently working as an assistant professor at IIT Hyderabad, India. He was the graduating topper of his batch in B.Tech and has received fellowship from ISU and performance award from ORNL. Sparsh has published more than 70 papers in top conferences and journals. His research interests include accelerators for machine learning, non-volatile memory, and GPU architectures. His webpage is http://www.iith.ac.in/~sparsh/

Subscribe to HPCwire's Weekly Update!

Be the most informed person in the room! Stay ahead of the tech trends with industy updates delivered to you every week!

Watch Nvidia’s GTC21 Keynote with Jensen Huang Livestreamed Here, Monday at 8:30am PT

April 9, 2021

Join HPCwire right here on Monday, April 12, at 8:30 am PT to see the Nvidia GTC21 keynote from Nvidia’s CEO, Jensen Huang, livestreamed in its entirety. Hosted by HPCwire, you can click to join the Huang keynote on our livestream to hear Nvidia’s expected news and... Read more…

The US Places Seven Additional Chinese Supercomputing Entities on Blacklist

April 8, 2021

As tensions between the U.S. and China continue to simmer, the U.S. government today added seven Chinese supercomputing entities to an economic blacklist. The U.S. Entity List bars U.S. firms from supplying key technolog Read more…

Argonne Supercomputing Supports Caterpillar Engine Design

April 8, 2021

Diesel fuels still account for nearly ten percent of all energy-related U.S. carbon emissions – most of them from heavy-duty vehicles like trucks and construction equipment. Energy efficiency is key to these machines, Read more…

Habana’s AI Silicon Comes to San Diego Supercomputer Center

April 8, 2021

Habana Labs, an Intel-owned AI company, has partnered with server maker Supermicro to provide high-performance, high-efficiency AI computing in the form of new training and inference servers that will power the upcoming Read more…

Intel Partners Debut Latest Servers Based on the New Intel Gen 3 ‘Ice Lake’ Xeons

April 7, 2021

Fresh from Intel’s launch of the company’s latest third-generation Xeon Scalable “Ice Lake” processors on April 6 (Tuesday), Intel server partners Cisco, Dell EMC, HPE and Lenovo simultaneously unveiled their first server models built around the latest chips. And though arch-rival AMD may... Read more…

AWS Solution Channel

Volkswagen Passenger Cars Uses NICE DCV for High-Performance 3D Remote Visualization

 

Volkswagen Passenger Cars has been one of the world’s largest car manufacturers for over 70 years. The company delivers more than 6 million automobiles to global customers every year, from 50 production locations on five continents. Read more…

What’s New in HPC Research: Tundra, Fugaku, µHPC & More

April 6, 2021

In this regular feature, HPCwire highlights newly published research in the high-performance computing community and related domains. From parallel programming to exascale to quantum computing, the details are here. Read more…

The US Places Seven Additional Chinese Supercomputing Entities on Blacklist

April 8, 2021

As tensions between the U.S. and China continue to simmer, the U.S. government today added seven Chinese supercomputing entities to an economic blacklist. The U Read more…

Habana’s AI Silicon Comes to San Diego Supercomputer Center

April 8, 2021

Habana Labs, an Intel-owned AI company, has partnered with server maker Supermicro to provide high-performance, high-efficiency AI computing in the form of new Read more…

Intel Partners Debut Latest Servers Based on the New Intel Gen 3 ‘Ice Lake’ Xeons

April 7, 2021

Fresh from Intel’s launch of the company’s latest third-generation Xeon Scalable “Ice Lake” processors on April 6 (Tuesday), Intel server partners Cisco, Dell EMC, HPE and Lenovo simultaneously unveiled their first server models built around the latest chips. And though arch-rival AMD may... Read more…

Intel Launches 10nm ‘Ice Lake’ Datacenter CPU with Up to 40 Cores

April 6, 2021

The wait is over. Today Intel officially launched its 10nm datacenter CPU, the third-generation Intel Xeon Scalable processor, codenamed Ice Lake. With up to 40 Read more…

HPE Launches Storage Line Loaded with IBM’s Spectrum Scale File System

April 6, 2021

HPE today launched a new family of storage solutions bundled with IBM’s Spectrum Scale Erasure Code Edition parallel file system (description below) and featu Read more…

RIKEN’s Ongoing COVID Research Includes New Vaccines, New Tests & More

April 6, 2021

RIKEN took the supercomputing world by storm last summer when it launched Fugaku – which became (and remains) the world’s most powerful supercomputer – ne Read more…

CERN Is Betting Big on Exascale

April 1, 2021

The European Organization for Nuclear Research (CERN) involves 23 countries, 15,000 researchers, billions of dollars a year, and the biggest machine in the worl Read more…

AI Systems Summit Keynote: Brace for System Level Heterogeneity Says de Supinski

April 1, 2021

Heterogeneous computing has quickly come to mean packing a couple of CPUs and one-or-many accelerators, mostly GPUs, onto the same node. Today, a one-such-node system has become the standard AI server offered by dozens of vendors. This is not to diminish the many advances... Read more…

Julia Update: Adoption Keeps Climbing; Is It a Python Challenger?

January 13, 2021

The rapid adoption of Julia, the open source, high level programing language with roots at MIT, shows no sign of slowing according to data from Julialang.org. I Read more…

Intel Launches 10nm ‘Ice Lake’ Datacenter CPU with Up to 40 Cores

April 6, 2021

The wait is over. Today Intel officially launched its 10nm datacenter CPU, the third-generation Intel Xeon Scalable processor, codenamed Ice Lake. With up to 40 Read more…

CERN Is Betting Big on Exascale

April 1, 2021

The European Organization for Nuclear Research (CERN) involves 23 countries, 15,000 researchers, billions of dollars a year, and the biggest machine in the worl Read more…

Programming the Soon-to-Be World’s Fastest Supercomputer, Frontier

January 5, 2021

What’s it like designing an app for the world’s fastest supercomputer, set to come online in the United States in 2021? The University of Delaware’s Sunita Chandrasekaran is leading an elite international team in just that task. Chandrasekaran, assistant professor of computer and information sciences, recently was named... Read more…

HPE Launches Storage Line Loaded with IBM’s Spectrum Scale File System

April 6, 2021

HPE today launched a new family of storage solutions bundled with IBM’s Spectrum Scale Erasure Code Edition parallel file system (description below) and featu Read more…

10nm, 7nm, 5nm…. Should the Chip Nanometer Metric Be Replaced?

June 1, 2020

The biggest cool factor in server chips is the nanometer. AMD beating Intel to a CPU built on a 7nm process node* – with 5nm and 3nm on the way – has been i Read more…

Saudi Aramco Unveils Dammam 7, Its New Top Ten Supercomputer

January 21, 2021

By revenue, oil and gas giant Saudi Aramco is one of the largest companies in the world, and it has historically employed commensurate amounts of supercomputing Read more…

Quantum Computer Start-up IonQ Plans IPO via SPAC

March 8, 2021

IonQ, a Maryland-based quantum computing start-up working with ion trap technology, plans to go public via a Special Purpose Acquisition Company (SPAC) merger a Read more…

Leading Solution Providers

Contributors

Can Deep Learning Replace Numerical Weather Prediction?

March 3, 2021

Numerical weather prediction (NWP) is a mainstay of supercomputing. Some of the first applications of the first supercomputers dealt with climate modeling, and Read more…

Livermore’s El Capitan Supercomputer to Debut HPE ‘Rabbit’ Near Node Local Storage

February 18, 2021

A near node local storage innovation called Rabbit factored heavily into Lawrence Livermore National Laboratory’s decision to select Cray’s proposal for its CORAL-2 machine, the lab’s first exascale-class supercomputer, El Capitan. Details of this new storage technology were revealed... Read more…

New Deep Learning Algorithm Solves Rubik’s Cube

July 25, 2018

Solving (and attempting to solve) Rubik’s Cube has delighted millions of puzzle lovers since 1974 when the cube was invented by Hungarian sculptor and archite Read more…

African Supercomputing Center Inaugurates ‘Toubkal,’ Most Powerful Supercomputer on the Continent

February 25, 2021

Historically, Africa hasn’t exactly been synonymous with supercomputing. There are only a handful of supercomputers on the continent, with few ranking on the Read more…

The History of Supercomputing vs. COVID-19

March 9, 2021

The COVID-19 pandemic poses a greater challenge to the high-performance computing community than any before. HPCwire's coverage of the supercomputing response t Read more…

HPE Names Justin Hotard New HPC Chief as Pete Ungaro Departs

March 2, 2021

HPE CEO Antonio Neri announced today (March 2, 2021) the appointment of Justin Hotard as general manager of HPC, mission critical solutions and labs, effective Read more…

AMD Launches Epyc ‘Milan’ with 19 SKUs for HPC, Enterprise and Hyperscale

March 15, 2021

At a virtual launch event held today (Monday), AMD revealed its third-generation Epyc “Milan” CPU lineup: a set of 19 SKUs -- including the flagship 64-core, 280-watt 7763 part --  aimed at HPC, enterprise and cloud workloads. Notably, the third-gen Epyc Milan chips achieve 19 percent... Read more…

Microsoft, HPE Bringing AI, Edge, Cloud to Earth Orbit in Preparation for Mars Missions

February 12, 2021

The International Space Station will soon get a delivery of powerful AI, edge and cloud computing tools from HPE and Microsoft Azure to expand technology experi Read more…

  • arrow
  • Click Here for More Headlines
  • arrow
HPCwire