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!

When in Rome: AMD Announces New Epyc CPU for HPC, Server and Cloud Wins

September 18, 2019

Where else but Rome could AMD hold the official Europe launch party for its second generation of Epyc microprocessors, codenamed Rome. Today, AMD did just that announcing key server wins, important cloud provider wins Read more…

By John Russell

Dell’s AMD-Powered Server Line Targets High-End Jobs

September 17, 2019

Dell Technologies rolled out five new servers this week based on AMD’s latest Epyc processor that are geared toward data-driven workloads running on increasingly popular multi-cloud platforms as well as in the HPC data Read more…

By George Leopold

Cerebras to Supply DOE with Wafer-Scale AI Supercomputing Technology

September 17, 2019

Cerebras Systems, which debuted its wafer-scale AI silicon at Hot Chips last month, has entered into a multi-year partnership with Argonne National Laboratory and Lawrence Livermore National Laboratory as part of a larger collaboration with the U.S. Department of Energy... Read more…

By Tiffany Trader

AWS Solution Channel

A Guide to Discovering the Best AWS Instances and Configurations for Your HPC Workload

The flexibility and heterogeneity of HPC cloud services provide a welcome contrast to the constraints of on-premises HPC. Every HPC configuration is potentially accessible to any given workload in a well-resourced cloud HPC deployment, with vast scalability to spin up as much compute as that workload demands in any given moment. Read more…

HPE Extreme Performance Solutions

Intel FPGAs: More Than Just an Accelerator Card

FPGA (Field Programmable Gate Array) acceleration cards are not new, as they’ve been commercially available since 1984. Typically, the emphasis around FPGAs has centered on the fact that they’re programmable accelerators, and that they can truly offer workload specific hardware acceleration solutions without requiring custom silicon. Read more…

IBM Accelerated Insights

Rumors of My Death Are Still Exaggerated: The Mainframe

[Connect with Spectrum users and learn new skills in the IBM Spectrum LSF User Community.]

As of 2017, 92 of the world’s top 100 banks used mainframes. Read more…

Better Scientific Software: Turn Your Passion into Cash

September 13, 2019

Do you know your way around scientific software and programming? You think you can contribute to the community by making scientific software better? If so, then the Better Scientific Software (BSSW) organization wants yo Read more…

By Dan Olds

When in Rome: AMD Announces New Epyc CPU for HPC, Server and Cloud Wins

September 18, 2019

Where else but Rome could AMD hold the official Europe launch party for its second generation of Epyc microprocessors, codenamed Rome. Today, AMD did just that Read more…

By John Russell

Cerebras to Supply DOE with Wafer-Scale AI Supercomputing Technology

September 17, 2019

Cerebras Systems, which debuted its wafer-scale AI silicon at Hot Chips last month, has entered into a multi-year partnership with Argonne National Laboratory and Lawrence Livermore National Laboratory as part of a larger collaboration with the U.S. Department of Energy... Read more…

By Tiffany Trader

IDAS: ‘Automagic’ HPC With Training Wheels

September 12, 2019

High-performance computing (HPC) for research is notorious for having steep barriers to entry. For this reason, high-tech disciplines were early adopters, have Read more…

By Elizabeth Leake

Univa Brings Cloud Automation to Slurm Users with Navops Launch 2.0

September 11, 2019

Univa, the company behind Grid Engine, announced today its HPC cloud-automation platform NavOps Launch will support the popular open-source workload scheduler Slurm. With the release of NavOps Launch 2.0, “Slurm users will have access to the same cloud automation capabilities... Read more…

By Tiffany Trader

When Dense Matrix Representations Beat Sparse

September 9, 2019

In our world filled with unintended consequences, it turns out that saving memory space to help deal with GPU limitations, knowing it introduces performance pen Read more…

By James Reinders

Eyes on the Prize: TACC’s Frontera Quickly Ramps up Science Agenda

September 9, 2019

Announced a year ago and officially launched a week ago, the Texas Advanced Computing Center’s Frontera – now the fastest academic supercomputer (~25 petefl Read more…

By John Russell

Quantum Roundup: IBM Goes to School, Delft Tackles Networking, Rigetti Updates

September 5, 2019

IBM today announced a new open source quantum ‘textbook’, a series of quantum education videos, and plans to expand its nascent quantum hackathon program. L Read more…

By John Russell

DARPA Looks to Propel Parallelism

September 4, 2019

As Moore’s law runs out of steam, new programming approaches are being pursued with the goal of greater hardware performance with less coding. The Defense Advanced Projects Research Agency is launching a new programming effort aimed at leveraging the benefits of massive distributed parallelism with less sweat. Read more…

By George Leopold

High Performance (Potato) Chips

May 5, 2006

In this article, we focus on how Procter & Gamble is using high performance computing to create some common, everyday supermarket products. Tom Lange, a 27-year veteran of the company, tells us how P&G models products, processes and production systems for the betterment of consumer package goods. Read more…

By Michael Feldman

Supercomputer-Powered AI Tackles a Key Fusion Energy Challenge

August 7, 2019

Fusion energy is the Holy Grail of the energy world: low-radioactivity, low-waste, zero-carbon, high-output nuclear power that can run on hydrogen or lithium. T Read more…

By Oliver Peckham

AMD Verifies Its Largest 7nm Chip Design in Ten Hours

June 5, 2019

AMD announced last week that its engineers had successfully executed the first physical verification of its largest 7nm chip design – in just ten hours. The AMD Radeon Instinct Vega20 – which boasts 13.2 billion transistors – was tested using a TSMC-certified Calibre nmDRC software platform from Mentor. Read more…

By Oliver Peckham

TSMC and Samsung Moving to 5nm; Whither Moore’s Law?

June 12, 2019

With reports that Taiwan Semiconductor Manufacturing Co. (TMSC) and Samsung are moving quickly to 5nm manufacturing, it’s a good time to again ponder whither goes the venerable Moore’s law. Shrinking feature size has of course been the primary hallmark of achieving Moore’s law... Read more…

By John Russell

DARPA Looks to Propel Parallelism

September 4, 2019

As Moore’s law runs out of steam, new programming approaches are being pursued with the goal of greater hardware performance with less coding. The Defense Advanced Projects Research Agency is launching a new programming effort aimed at leveraging the benefits of massive distributed parallelism with less sweat. Read more…

By George Leopold

Cray Wins NNSA-Livermore ‘El Capitan’ Exascale Contract

August 13, 2019

Cray has won the bid to build the first exascale supercomputer for the National Nuclear Security Administration (NNSA) and Lawrence Livermore National Laborator Read more…

By Tiffany Trader

AMD Launches Epyc Rome, First 7nm CPU

August 8, 2019

From a gala event at the Palace of Fine Arts in San Francisco yesterday (Aug. 7), AMD launched its second-generation Epyc Rome x86 chips, based on its 7nm proce Read more…

By Tiffany Trader

Nvidia Embraces Arm, Declares Intent to Accelerate All CPU Architectures

June 17, 2019

As the Top500 list was being announced at ISC in Frankfurt today with an upgraded petascale Arm supercomputer in the top third of the list, Nvidia announced its Read more…

By Tiffany Trader

Leading Solution Providers

ISC 2019 Virtual Booth Video Tour

CRAY
CRAY
DDN
DDN
DELL EMC
DELL EMC
GOOGLE
GOOGLE
ONE STOP SYSTEMS
ONE STOP SYSTEMS
PANASAS
PANASAS
VERNE GLOBAL
VERNE GLOBAL

Ayar Labs to Demo Photonics Chiplet in FPGA Package at Hot Chips

August 19, 2019

Silicon startup Ayar Labs continues to gain momentum with its DARPA-backed optical chiplet technology that puts advanced electronics and optics on the same chip Read more…

By Tiffany Trader

Top500 Purely Petaflops; US Maintains Performance Lead

June 17, 2019

With the kick-off of the International Supercomputing Conference (ISC) in Frankfurt this morning, the 53rd Top500 list made its debut, and this one's for petafl Read more…

By Tiffany Trader

A Behind-the-Scenes Look at the Hardware That Powered the Black Hole Image

June 24, 2019

Two months ago, the first-ever image of a black hole took the internet by storm. A team of scientists took years to produce and verify the striking image – an Read more…

By Oliver Peckham

Cray – and the Cray Brand – to Be Positioned at Tip of HPE’s HPC Spear

May 22, 2019

More so than with most acquisitions of this kind, HPE’s purchase of Cray for $1.3 billion, announced last week, seems to have elements of that overused, often Read more…

By Doug Black and Tiffany Trader

Chinese Company Sugon Placed on US ‘Entity List’ After Strong Showing at International Supercomputing Conference

June 26, 2019

After more than a decade of advancing its supercomputing prowess, operating the world’s most powerful supercomputer from June 2013 to June 2018, China is keep Read more…

By Tiffany Trader

Qualcomm Invests in RISC-V Startup SiFive

June 7, 2019

Investors are zeroing in on the open standard RISC-V instruction set architecture and the processor intellectual property being developed by a batch of high-flying chip startups. Last fall, Esperanto Technologies announced a $58 million funding round. Read more…

By George Leopold

Intel Confirms Retreat on Omni-Path

August 1, 2019

Intel Corp.’s plans to make a big splash in the network fabric market for linking HPC and other workloads has apparently belly-flopped. The chipmaker confirmed to us the outlines of an earlier report by the website CRN that it has jettisoned plans for a second-generation version of its Omni-Path interconnect... Read more…

By Staff report

Intel Debuts Pohoiki Beach, Its 8M Neuron Neuromorphic Development System

July 17, 2019

Neuromorphic computing has received less fanfare of late than quantum computing whose mystery has captured public attention and which seems to have generated mo Read more…

By John Russell

  • arrow
  • Click Here for More Headlines
  • arrow
Do NOT follow this link or you will be banned from the site!
Share This