So, what’s next? A year after computers formally crossed the “exascale” milestone – becoming capable of crunching a quintillion (1018) calculations per second – what’s next on the horizon? What’s the next supercomputing initiative? And now that we’ve officially shifted eras, from petascale to “post-exascale,” what does that mean for scientists, for tech users and for society at large?
Researchers at Hewlett Packard Labs are hard at work, seeking answers to these and other questions about where high-performance computing (HPC) is headed. They’re paving the way to a post-exascale future by studying and trying to solve the real, pressing logistical issues that the industry faces today. With the end of Dennard scaling and the slowing down of Moore’s Law, how can you architect a supercomputer that is 10 times faster than the first exascale system? Can you build such a system in the next few years? How do you design supercomputers for emerging customer compute needs with data science and artificial intelligence workflows? How do you make exascale computing accessible and consumable from the edge and in the cloud? How do you build systems that run the world’s fastest workflows?
Supercomputers certainly have come a long way. They’re already running everyday applications and remarkable speeds and feeds. Industries such as oil exploration, finance, personalized content-delivery and recommendation systems, and online advertising are deploying high-performance computing (HPC) systems to manage heavy workloads delivering real-time services.
AI has arrived
The biggest change in the exascale and post-exascale eras is that artificial intelligence (AI) has arrived – in a big way. As industries move to incorporate more AI, they’re crunching massive amounts of data to teach the systems how to work. Coupling HPC with AI allows industries to train better, bigger, and more accurate models, and the workloads to train AI models are as exascale hungry as the physics models with traditional HPC.
The AI-HPC combination is delivering a multiplier effect in all kinds of modern applications. Self-driving cars use AI models trained on supercomputers. Economics researchers are applying AI models to satellite images of retail parking lots to refine revenue projections. And drug makers are using generative AI models to synthesize molecules with desired treatment properties, saving time and money, while being more accurate. At least two companies, in fact, have already introduced drugs into clinical trials that have been designed fully using AI.
All of this requires exascale-level horsepower. In the 13 years since computers reached petascale for the first time (1015 calculations per second), storage capacity on the world’s fastest supercomputers have increased by a factor of 70 and bandwidths available to move data have grown by a factor of 20. Models trained/computed on the supercomputers are over 8,000 times larger in size, and there are 100 times as many new models to crunch. Consumption of compute is growing across the board. Today’s multifaceted, complicated workflows with artificial intelligence infused are consuming 300,000 times more compute.
"Now that we're a year into the Exascale Era, two things are clear: Supercomputers are growing much more powerful, and they’re needed more than ever," said Hewlett Packard Labs Director Andrew Wheeler, who is leading an exascale discussion session June 21 at HPE Discover. "But will the computing function evolve to the point where it can continue to optimize tomorrow’s use cases? That will depend on how well researchers can pivot their design principles to support innovations on three fronts – systems, workflows, and operational efficiencies."
Here is where Hewlett Packard Labs is focusing its energies.
Systems
The capabilities of supercomputing systems will have to evolve in several ways. As Moore’s Law slows, customers expect HPE servers to live longer in the data center. System architects will need to create systems that can accommodate more application-specific processor architectures. These specialized architectures will have to work seamlessly over a network designed with high-performance interconnect capabilities that deliver speed and bandwidth but is also dynamically configurable because some workloads perform better on certain topologies.
Workflows
Delivering performance for end-to-end workflows comes with its own challenges. In the post-exascale future, workflows (that can be defined as a semi-automated complex sequence of workloads) will touch a spectrum of resources across the whole computing landscape, from the edge to on-prem supercomputers to the cloud. For these workflows, data can emerge from an edge device, from a simulation/digital twin or from a massive data store. Data will have to be managed in all its forms – at rest, triggered by events, and in motion. Workloads will need to be orchestrated across mediums, and analytics will need to be managed to serve persistent, serverless and batch processing scenarios. In the post-exascale era, users need to be able to orchestrate workloads that connect edge instruments of interest for any industry – be it telecom, retail, manufacturing, or cameras installed for security/safety purposes. HPE systems in the post-exascale era will provide different knobs of performance (capacity, capability, latency, and scalability) to these emerging workflows. That is, given any workflow, a customer can build a supercomputer that will run that specific workflow the fastest.
Operational Efficiencies
In the post-exascale world, design, implementation, and optimization of workflows on supercomputing systems will demonstrate HPE’s approach toward operational efficiency on several levels. New supercomputing platforms will need to secure infrastructure to customers’ desired need for security/privacy. They’ll need to track and reveal opportunities for efficiency in support of sustainability goals of energy consumption. And they’ll need to deliver flexible financial and consumption models of compute resources and services.
Conclusion
In summary, exascale computing excellence was engineered performance by design – posing and addressing the challenge of building supercomputers within power and cost budgets to perform a quintillion math operations per second. In this new post-Exascale Era of converged HPC and AI, where “workflows are the new applications,” building supercomputers for end-to-end workflows is going to be about engineering capability that is dynamic to (i) availability/accessibility of resources at the time of execution, (ii) efficiency (both code portability and energy), and (iii) configurability for performance. This HPE approach enables architectural creativity to deliver flexible consumption models of exaflop-seconds for analytics, exaflop-hours for HPC codes, and exaflop-months for AI models.
To learn more about the post-exascale era of converged HPC and AI, attend Andrew Wheeler’s session June 21 at HPE Discover.
