Accelerating AI Transformation by Making HPC Easier

Figen Ülgen, PhD, is General Manager, Rack Scale Design for Intel's Data Center Group.

Why is someone whose doctoral thesis dealt with neural network training focusing now on data center innovation?

The short answer: Because artificial intelligence (AI) changes everything.

AI is already enriching our daily lives. We interact with AI as we use our digital assistants, search for information on the web, or are offered shopping recommendations. From manufacturing to finance, AI is adding intelligence, supporting human expertise, and promoting smarter decision making. And we only at the early stages.

On the farm and in the greenhouse, solutions that combine AI with the Internet of Things (IoT) are optimizing the delivery of water and nutrients. As they mature and are applied more broadly, these technologies have the potential to increase crop yields and minimize resource consumption, to help feed a growing global population on a changing planet.  

On the road, AI is delivering driver-assistance features that augment human awareness and responses. As innovators make further progress, autonomous vehicles offer potential breakthroughs in fuel usage, air quality, efficiency, safety, as well as in human productivity, enjoyment, and more.

In healthcare, AI-based medical-imaging solutions are showing they can examine x-rays and CT scans and identify malignant tumors more quickly and accurately than highly trained experts. Incorporated into radiology workflows, these AI solutions can help governments, health systems, and clinicians deliver high-quality healthcare and reduce clinician burnout despite rising demand for care, worldwide staffing shortages, and constrained budgets. And medical imaging is a single area. Healthcare providers, payers, pharma researchers, and other stakeholders are exploring and implementing dozens of use cases where they expect AI to help them improve population health and better meet both business and clinical objectives.

In the data center, AI will bring unprecedented optimization to rack-level and data-center-wide design on top of hardware advances in data center design. That’s one area where I’m currently applying my long-standing passion for intelligent systems.

More is Needed

As exciting as these developments are, however, we need to do more if AI is to reach its full potential. We must make it easier for AI analysts and domain experts to take full advantage of underlying platforms such as high-performance computing (HPC) and other relevant hardware innovation across the data center.

I believe all of us have a role in making that happen. And given the powerful benefits to be achieved and problems to be solved, we all have a stake in making the AI/HPC convergence happen now.                                                                                                                                                        

Model Training Demands HPC

Deep learning and other branches of AI are a natural fit for HPC. For example, deep or convolutional neural network algorithms, which have been my focus, can be trained to identify and classify objects (is that object a stop sign or a tree? A malignancy or normal tissue?). Used with other algorithms, deep learning models can also recommend appropriate responses. Apply brakes or proceed ahead? Flag for immediate follow-up, recommend continued monitoring, or mark as normal?

Neural networks represent a vast improvement over the more traditional, rules-based approaches to AI that dominated my earlier career. Rather than needing developers to explicitly define the relevant rules and features that govern an AI model, neural network algorithms can churn through vast amounts of relevant data and infer the rules, which developers then confirm.

That’s the good news. However, the work of training a deep learning model to recognize objects and derive rules and features requires feeding the model large numbers of examples—a task that is both compute and data intensive. For example, training a decent speech recognition model, a task common to many AI solutions, requires processing thousands of hours of speech data. The more computing performance you apply to training a model, other things being equal, the faster you can get a working model that provides reliable answers. The more data you use to train the model, the higher level of accuracy it can achieve and the more confidence you can have in its conclusions.

Advances in computing technologies have transformed turnaround times since the days when my neural network training algorithms, coupled with natural language processing, ran for days. In addition, many aspects of model training are highly parallelizable, enabling AI developers to speed model training further by subdividing their training tasks and computing them simultaneously on massively parallel HPC systems. Using HPC and parallel processing for model training, AI developers can obtain more accurate working models more quickly. Developers can also create models capable of tackling more complex inferences and solving harder problems with more and deeper layers in the neural networks.

These benefits are helping make AI the fastest-growing segment of the global HPC server market. HPC’s AI sector is in the midst of a five-year cumulative annual growth rate (CAGR) of 29.5 percent, rising to $1.3 billion in 2021.

But for many domain experts and data scientists, HPC is out of reach. In part 2 of this article, I’ll describe an open source initiative that makes it easier for business and research projects to leverage the power of HPC—and deliver on the vision of AI everywhere.

Opinions expressed in the article above do not necessarily reflect the opinions of Data Center Knowledge and Informa.

Industry Perspectives is a content channel at Data Center Knowledge highlighting thought leadership in the data center arena. See our guidelines and submission process for information on participating.