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

Subscribe to HPCwire's Weekly Update!

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

Quantum Watchers – Terrific Interview with Caltech’s John Preskill by CERN

July 17, 2024

In case you missed it, there's a fascinating interview with John Preskill, the prominent Caltech physicist and pioneering quantum computing researcher that was recently posted by CERN’s department of experimental physi Read more…

Aurora AI-Driven Atmosphere Model is 5,000x Faster Than Traditional Systems

July 16, 2024

While the onset of human-driven climate change brings with it many horrors, the increase in the frequency and strength of storms poses an enormous threat to communities across the globe. As climate change is warming ocea Read more…

Researchers Say Memory Bandwidth and NVLink Speeds in Hopper Not So Simple

July 15, 2024

Researchers measured the real-world bandwidth of Nvidia's Grace Hopper superchip, with the chip-to-chip interconnect results falling well short of theoretical claims. A paper published on July 10 by researchers in the U. Read more…

Belt-Tightening in Store for Most Federal FY25 Science Budets

July 15, 2024

If it’s summer, it’s federal budgeting time, not to mention an election year as well. There’s an excellent summary of the curent state of FY25 efforts reported in AIP’s policy FYI: Science Policy News. Belt-tight Read more…

Peter Shor Wins IEEE 2025 Shannon Award

July 15, 2024

Peter Shor, the MIT mathematician whose ‘Shor’s algorithm’ sent shivers of fear through the encryption community and helped galvanize ongoing efforts to build quantum computers, has been named the 2025 winner of th Read more…

Weekly Wire Roundup: July 8-July 12, 2024

July 12, 2024

HPC news can get pretty sleepy in June and July, but this week saw a bump in activity midweek as Americans realized they still had work to do after the previous holiday weekend. The world outside the United States also s Read more…

Aurora AI-Driven Atmosphere Model is 5,000x Faster Than Traditional Systems

July 16, 2024

While the onset of human-driven climate change brings with it many horrors, the increase in the frequency and strength of storms poses an enormous threat to com Read more…

Shutterstock 1886124835

Researchers Say Memory Bandwidth and NVLink Speeds in Hopper Not So Simple

July 15, 2024

Researchers measured the real-world bandwidth of Nvidia's Grace Hopper superchip, with the chip-to-chip interconnect results falling well short of theoretical c Read more…

Shutterstock 2203611339

NSF Issues Next Solicitation and More Detail on National Quantum Virtual Laboratory

July 10, 2024

After percolating for roughly a year, NSF has issued the next solicitation for the National Quantum Virtual Lab program — this one focused on design and imple Read more…

NCSA’s SEAS Team Keeps APACE of AlphaFold2

July 9, 2024

High-performance computing (HPC) can often be challenging for researchers to use because it requires expertise in working with large datasets, scaling the softw Read more…

Anders Jensen on Europe’s Plan for AI-optimized Supercomputers, Welcoming the UK, and More

July 8, 2024

The recent ISC24 conference in Hamburg showcased LUMI and other leadership-class supercomputers co-funded by the EuroHPC Joint Undertaking (JU), including three Read more…

Generative AI to Account for 1.5% of World’s Power Consumption by 2029

July 8, 2024

Generative AI will take on a larger chunk of the world's power consumption to keep up with the hefty hardware requirements to run applications. "AI chips repres Read more…

US Senators Propose $32 Billion in Annual AI Spending, but Critics Remain Unconvinced

July 5, 2024

Senate leader, Chuck Schumer, and three colleagues want the US government to spend at least $32 billion annually by 2026 for non-defense related AI systems.  T Read more…

Point and Click HPC: High-Performance Desktops

July 3, 2024

Recently, an interesting paper appeared on Arvix called Use Cases for High-Performance Research Desktops. To be clear, the term desktop in this context does not Read more…

Atos Outlines Plans to Get Acquired, and a Path Forward

May 21, 2024

Atos – via its subsidiary Eviden – is the second major supercomputer maker outside of HPE, while others have largely dropped out. The lack of integrators and Atos' financial turmoil have the HPC market worried. If Atos goes under, HPE will be the only major option for building large-scale systems. Read more…

Everyone Except Nvidia Forms Ultra Accelerator Link (UALink) Consortium

May 30, 2024

Consider the GPU. An island of SIMD greatness that makes light work of matrix math. Originally designed to rapidly paint dots on a computer monitor, it was then Read more…

Comparing NVIDIA A100 and NVIDIA L40S: Which GPU is Ideal for AI and Graphics-Intensive Workloads?

October 30, 2023

With long lead times for the NVIDIA H100 and A100 GPUs, many organizations are looking at the new NVIDIA L40S GPU, which it’s a new GPU optimized for AI and g Read more…


Nvidia Economics: Make $5-$7 for Every $1 Spent on GPUs

June 30, 2024

Nvidia is saying that companies could make $5 to $7 for every $1 invested in GPUs over a four-year period. Customers are investing billions in new Nvidia hardwa Read more…

Nvidia Shipped 3.76 Million Data-center GPUs in 2023, According to Study

June 10, 2024

Nvidia had an explosive 2023 in data-center GPU shipments, which totaled roughly 3.76 million units, according to a study conducted by semiconductor analyst fir Read more…

AMD Clears Up Messy GPU Roadmap, Upgrades Chips Annually

June 3, 2024

In the world of AI, there's a desperate search for an alternative to Nvidia's GPUs, and AMD is stepping up to the plate. AMD detailed its updated GPU roadmap, w Read more…

Some Reasons Why Aurora Didn’t Take First Place in the Top500 List

May 15, 2024

The makers of the Aurora supercomputer, which is housed at the Argonne National Laboratory, gave some reasons why the system didn't make the top spot on the Top Read more…

Intel’s Next-gen Falcon Shores Coming Out in Late 2025 

April 30, 2024

It's a long wait for customers hanging on for Intel's next-generation GPU, Falcon Shores, which will be released in late 2025.  "Then we have a rich, a very Read more…

Leading Solution Providers


Google Announces Sixth-generation AI Chip, a TPU Called Trillium

May 17, 2024

On Tuesday May 14th, Google announced its sixth-generation TPU (tensor processing unit) called Trillium.  The chip, essentially a TPU v6, is the company's l Read more…

Nvidia H100: Are 550,000 GPUs Enough for This Year?

August 17, 2023

The GPU Squeeze continues to place a premium on Nvidia H100 GPUs. In a recent Financial Times article, Nvidia reports that it expects to ship 550,000 of its lat Read more…

IonQ Plots Path to Commercial (Quantum) Advantage

July 2, 2024

IonQ, the trapped ion quantum computing specialist, delivered a progress report last week firming up 2024/25 product goals and reviewing its technology roadmap. Read more…

Choosing the Right GPU for LLM Inference and Training

December 11, 2023

Accelerating the training and inference processes of deep learning models is crucial for unleashing their true potential and NVIDIA GPUs have emerged as a game- Read more…

Nvidia’s New Blackwell GPU Can Train AI Models with Trillions of Parameters

March 18, 2024

Nvidia's latest and fastest GPU, codenamed Blackwell, is here and will underpin the company's AI plans this year. The chip offers performance improvements from Read more…

The NASA Black Hole Plunge

May 7, 2024

We have all thought about it. No one has done it, but now, thanks to HPC, we see what it looks like. Hold on to your feet because NASA has released videos of wh Read more…

Q&A with Nvidia’s Chief of DGX Systems on the DGX-GB200 Rack-scale System

March 27, 2024

Pictures of Nvidia's new flagship mega-server, the DGX GB200, on the GTC show floor got favorable reactions on social media for the sheer amount of computing po Read more…

MLPerf Inference 4.0 Results Showcase GenAI; Nvidia Still Dominates

March 28, 2024

There were no startling surprises in the latest MLPerf Inference benchmark (4.0) results released yesterday. Two new workloads — Llama 2 and Stable Diffusion Read more…

  • arrow
  • Click Here for More Headlines
  • arrow