ORNL: World’s Fastest Supercomputer Takes on Large Language Modeling

May 14, 2024

May 14, 2024 — A team led by researchers at the Department of Energy’s Oak Ridge National Laboratory explored training strategies for one of the largest artificial intelligence models to date with help from the world’s fastest supercomputer.

The findings could help guide training for a new generation of AI models for scientific research.

The study led by ORNL’s Sajal Dash, Feiyi Wang and Prasanna Balaprakash employed Frontier, the world’s first exascale supercomputer, to run the initial stages of training on a large language model similar to OpenAI’s ChatGPT. The research team used a set of test data to project how models with 22 billion, 175 billion, and 1 trillion parameters, or variables, could run across 128 and later 384 of Frontier’s more than 9,400 nodes. The team didn’t attempt to train a full model to completion.

“This study and our findings aren’t so much a manual as a potential set of guidelines for users training a large model,” Dash said. “They can draw from our experience to decide how to use Frontier’s resources to train their particular model and make the most effective use of their allotted computing time.”

The team will present the study at the International Supercomputing Conference High Performance 2024 in May in Hamburg, Germany. Fellow scientists Isaac Lyngaas, Junqi Yin, Xiao Wang and Guojing Cong of ORNL and Romaine Egele of Paris-Saclay University also collaborated on the study.

The study focused less on model development than on pinpointing the most efficient ways to exploit the graphics processing units, or GPUs, that power Frontier and similar supercomputers and putting them to work training an AI. Each of Frontier’s nodes relies on four AMD MI250X GPUs for a total of more than 75,000 GPUs.

The training ran for a few hours on about 100 million tokens — basic units of text such as words and characters — of test data. That’s about a ten-thousandth of the necessary data to train a trillion-parameter model to completion and an even smaller fraction of the necessary time.

The research team used the data from those runs to calculate how a trillion-parameter model might perform if trained to completion on Frontier.

“This study was largely an exercise to show we can train this particular size of model on Frontier at this particular scale with this particular level of efficiency,” Wang said. “We didn’t get anywhere near the finish line of a complete large language model.”

Large language models loosely mimic the human brain in their ability to learn and recognize patterns in words and numbers and to improve on that learning over time with additional training. The goal: design a model that can absorb and adjust the lessons learned on training data and apply that knowledge consistently and accurately to new, unfamiliar data and tasks.

The vast datasets and powerful processors needed for such training have remained mostly out of reach of scholars and in the possession of private companies, which tend to guard those resources as proprietary and set strict conditions for use. Those conditions typically limit research opportunities and don’t allow results to be easily verified. But leadership-class supercomputers like Frontier, which awards computing time to scientific researchers through the DOE’s Innovative and Novel Computational Impact on Theory and Experiment program, could enable a new generation of AI models to be trained more quickly if scientists find the right approach.

“Traditionally, this process has relied on expert knowledge or on trial and error,” said Prasanna Balaprakash, ORNL’s director of AI programs. “One of the highlights of our work in this study is the automation of identifying high-performing strategies among a vast array of options. We leveraged DeepHyper, an open-source scalable tuning software, to automatically determine the optimal settings. We plan to extend this automated approach to fine-tune system-level performance and enhance efficiency at an extreme scale. Furthermore, we have democratized our methodologies and software for the benefit of the scientific community. This strategy ensures that our insights are widely accessible for future research on training large AI foundation models in science.”

The larger the model and its training datasets, the better its performance — but also the higher its demand for computational power. Training a trillion-parameter large language model from the initial stages to completion without optimizations would take months even at Frontier’s world-leading speeds.

The ORNL study examined approaches to data parallelism — a process used by supercomputers like Frontier to break a large problem into smaller problems to reach a solution more quickly — to train AI and how to port that training across proprietary frameworks of GPUs made by competing vendors.

“It’s about finding the best combination of training strategies while getting the best throughput,” Dash said. “Most deep-learning frameworks target the GPUs made by NVIDIA rather than the GPUs made by AMD that power Frontier. We wanted to see if existing models could run on Frontier, how to make the best use of Frontier’s computing power and how to make that level of performance possible across GPU platforms.

“We can’t train a model this size on a single GPU or a single node, for example, and every time we cross the barrier between nodes that requires more communication that consumes more time. How do we slice up the model across GPUs so that we can fit and train the model without losing too much time and energy communicating between nodes?”

The researchers found a blend of parallelism strategies worked best when tailored to the computing platform but said their work’s far from finished.

“The efficiency we achieved on Frontier with this model was decent but not decent enough,” Wang said. “At extreme scale, we achieved 30% efficiency — which means we left about 70% of Frontier’s computing power on the floor. We need much more optimization to make the machine more efficient at this scale.”

The team’s next steps include training a model further with peer-reviewed scientific data across more nodes.

Support for this research came from the DOE Office of Science’s Advanced Scientific Computing Research program and ORNL’s AI Initiative. The OLCF is a DOE Office of Science user facility.

UT-Battelle manages ORNL for DOE’s Office of Science, the single largest supporter of basic research in the physical sciences in the United States. DOE’s Office of Science is working to address some of the most pressing challenges of our time. For more information, visit https://energy.gov/science.


Source: Oak Ridge National Laboratory

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!

ISC 2024 Takeaways: Love for Top500, Extending HPC Systems, and Media Bashing

May 23, 2024

The ISC High Performance show is typically about time-to-science, but breakout sessions also focused on Europe's tech sovereignty, server infrastructure, storage, throughput, and new computing technologies. This round Read more…

HPC Pioneer Gordon Bell Passed Away

May 22, 2024

Legendary computer scientist Gordon Bell passed away last Friday at his home in Coronado, CA. He was 89. The New York Times has a nice tribute piece. A long-time pioneer with Digital Equipment Corp, he pushed hard for de Read more…

ISC 2024 — A Few Quantum Gems and Slides from a Packed QC Agenda

May 22, 2024

If you were looking for quantum computing content, ISC 2024 was a good place to be last week — there were around 20 quantum computing related sessions. QC even earned a slide in Kathy Yelick’s opening keynote — Bey 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…

Core42 Is Building Its 172 Million-core AI Supercomputer in Texas

May 20, 2024

UAE-based Core42 is building an AI supercomputer with 172 million cores which will become operational later this year. The system, Condor Galaxy 3, was announced earlier this year and will have 192 nodes with Cerebras Read more…

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 latest weapon in the AI battle with GPU maker Nvidia and clou Read more…

ISC 2024 Takeaways: Love for Top500, Extending HPC Systems, and Media Bashing

May 23, 2024

The ISC High Performance show is typically about time-to-science, but breakout sessions also focused on Europe's tech sovereignty, server infrastructure, storag Read more…

ISC 2024 — A Few Quantum Gems and Slides from a Packed QC Agenda

May 22, 2024

If you were looking for quantum computing content, ISC 2024 was a good place to be last week — there were around 20 quantum computing related sessions. QC eve 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…

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…

Europe’s Race towards Quantum-HPC Integration and Quantum Advantage

May 16, 2024

What an interesting panel, Quantum Advantage — Where are We and What is Needed? While the panelists looked slightly weary — their’s was, after all, one of Read more…

The Future of AI in Science

May 15, 2024

AI is one of the most transformative and valuable scientific tools ever developed. By harnessing vast amounts of data and computational power, AI systems can un 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…

ISC 2024 Keynote: High-precision Computing Will Be a Foundation for AI Models

May 15, 2024

Some scientific computing applications cannot sacrifice accuracy and will always require high-precision computing. Therefore, conventional high-performance c Read more…

Synopsys Eats Ansys: Does HPC Get Indigestion?

February 8, 2024

Recently, it was announced that Synopsys is buying HPC tool developer Ansys. Started in Pittsburgh, Pa., in 1970 as Swanson Analysis Systems, Inc. (SASI) by John Swanson (and eventually renamed), Ansys serves the CAE (Computer Aided Engineering)/multiphysics engineering simulation market. 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…

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…

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…

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…

AMD MI3000A

How AMD May Get Across the CUDA Moat

October 5, 2023

When discussing GenAI, the term "GPU" almost always enters the conversation and the topic often moves toward performance and access. Interestingly, the word "GPU" is assumed to mean "Nvidia" products. (As an aside, the popular Nvidia hardware used in GenAI are not technically... 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…

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…

Leading Solution Providers

Contributors

Shutterstock 1606064203

Meta’s Zuckerberg Puts Its AI Future in the Hands of 600,000 GPUs

January 25, 2024

In under two minutes, Meta's CEO, Mark Zuckerberg, laid out the company's AI plans, which included a plan to build an artificial intelligence system with the eq Read more…

Eyes on the Quantum Prize – D-Wave Says its Time is Now

January 30, 2024

Early quantum computing pioneer D-Wave again asserted – that at least for D-Wave – the commercial quantum era has begun. Speaking at its first in-person Ana Read more…

The GenAI Datacenter Squeeze Is Here

February 1, 2024

The immediate effect of the GenAI GPU Squeeze was to reduce availability, either direct purchase or cloud access, increase cost, and push demand through the roof. A secondary issue has been developing over the last several years. Even though your organization secured several racks... Read more…

Shutterstock 1285747942

AMD’s Horsepower-packed MI300X GPU Beats Nvidia’s Upcoming H200

December 7, 2023

AMD and Nvidia are locked in an AI performance battle – much like the gaming GPU performance clash the companies have waged for decades. AMD has claimed it 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…

Intel Plans Falcon Shores 2 GPU Supercomputing Chip for 2026  

August 8, 2023

Intel is planning to onboard a new version of the Falcon Shores chip in 2026, which is code-named Falcon Shores 2. The new product was announced by CEO Pat Gel Read more…

GenAI Having Major Impact on Data Culture, Survey Says

February 21, 2024

While 2023 was the year of GenAI, the adoption rates for GenAI did not match expectations. Most organizations are continuing to invest in GenAI but are yet to Read more…

How the Chip Industry is Helping a Battery Company

May 8, 2024

Chip companies, once seen as engineering pure plays, are now at the center of geopolitical intrigue. Chip manufacturing firms, especially TSMC and Intel, have b Read more…

  • arrow
  • Click Here for More Headlines
  • arrow
HPCwire