Pittsburgh Supercomputing Center (PSC – a joint research organization of Carnegie Mellon University and the University of Pittsburgh) has won a $5 million award from the National Science Foundation to build an AI supercomputer designed to accelerate AI research in pursuit of science, discovery and societal good. The new machine, called Neocortex, couples two Cerebras CS-1 AI servers with a shared-memory HPE Superdome Flex server. PSC will make Neocortex available to researchers across the Extreme Science and Engineering Discovery Environment (XSEDE) later this year.
Each Cerebras CS-1 is powered by one Cerebras Wafer Scale Engine (WSE) processor, which contains 400,000 AI-optimized cores implemented on a 46,225 square millimeter wafer with 1.2 trillion transistors. A front-end HPE Superdome Flex server will handle pre- and post-processing of data flowing in and out of the WSE processors. The HPE Superdome Flex is provisioned with 32 Intel Xeon CPUs, 24 terabytes of memory, 205 terabytes of flash storage, and 24 network interface cards.
The Superdome Flex connects to each CS-1 server via 12 100 gigabit Ethernet links, providing 1.2 terabits per second of bandwidth between the machines. That’s enough bandwidth to transfer 37 HD movies every second, said Nick Nystrom, chief scientist, Pittsburgh Supercomputing Center. The Neocortex team is considering implementing the network on a single switch to explore allowing the two CS-1s to interface directly at 1.2 terabits per second.
The WSE processor inside the CS-1 provides 9 petabytes per second of on-die memory bandwidth, equivalent to about a million HD movies per second, by Nystrom’s math.
Neocortex (named after the region of the brain responsible for higher-order brain functions, including language processing) is the first CS-1 installation funded by the NSF and the first publicly announced CS-1 cluster. Cerebras debuted its Wafer Scale Engine last August at Hot Chips and the CS-1 system unveiling followed at SC19 in November. The Department of Energy was the flagship customer; single-node CS-1 systems are deployed at Argonne National Lab and Lawrence Livermore National Lab.
Describing the impetus for the technology partnering, Nystrom said that PSC saw the opportunity to bring together the best of two worlds – “the extreme deep learning capability of the server CS-1, and the extreme shared memory of the Superdome Flex with HPE.”
“With shared memory, you don’t have to break your problem across many nodes. You don’t have to write MPI, and you don’t have to distribute your data structures. It’s just all there at high speed,” he added.
Both Cerebras and PSC expressed their expectation that the system will be able to take on a new class of problems, beyond what is available with traditional GPUs.
“We’re just scratching the surface of sort of a new class of AI models; we know of additional models that have been difficult to get running on graphics processing units and we are extremely eager to be partnering with pioneering researchers to show the world what these models might be able to do,” said Andrew Feldman, Cerebras cofounder and CEO. His list of target examples includes models with separable convolutions or models with native and induced sparsity, both coarse and fine grained, graph neural networks with irregular sparse connections, complex sequential models, and very large models where parallelism is desirable.
Even with current best-in-class PSC machines, like the GPU-based Bridges and Bridges-AI, research is constrained, said Paola Buitrago, principal investigator and PSC director of artificial intelligence and big data, noting “there is clearly a need for more compute, and fast interconnect and storage.”
“Artificial intelligence in 2012 started this kind of renaissance, thanks to neural networks being implemented on GPUs,” Buitrago shared in an interview with HPCwire. “GPUs absolutely do well with matrix operations, which is one of the main operations in our neural networks, but they weren’t designed for AI. Now with the Cerebras technology, we see a machine that is specifically designed for AI and for the potential optimizations in deep learning. We are excited to explore how it can speed up and transform what is currently happening in deep learning, allowing us to explore more and more ambitious science and reducing the time to curiosity.”
Buitrago expects Neocortex to be more powerful than the PSC Bridges-AI system by a few orders of magnitude. Providing further characterization of the system’s potential, Cerebras’ Feldman said the tuned system cluster with Cerebras’ wafer-scale cores and “the pre-processing machine from HPE” will have the power of 800-1,500 traditional GPUs, or “or about 20 racks worth of graphics processing with a single rack of Cerebras.”
Naturally, PSC will be putting Neocortex through its paces to see if this claim bears out. The Neocortex group at PSC has identified a number of benchmarks as being important to the community. “These were selected to demonstrate the capability of the system when it hits the ground, and the system will, of course, continue to mature over time,” said Nystrom, adding they will be evaluating the system with all the big complex networks that are very challenging right now, including LSTM.
“In addition to LSTM, we expect Neocortex will be very good at graph convolutional networks, important in all kinds of science,” said Nystrom. “And then over time across CNNs. So we’ll be using those initially, and we’ll be engaging early users to demonstrate scientific impact. That’s very important to the National Science Foundation.”
Buitrago said that their users who are bounded by current hardware are “in large part working on natural language processing and working with transformer type networks, including BERT and Megatron-LM, where the models are quite big with hundreds of millions and billions of parameters,” adding, “that’s a specific use case that we will be enabling with the Neocortex system.”
The number of applications that need AI is growing, encompassing virtually all fields of science, many drawing on computer vision, text processing, and natural language processing. “We want to explore use cases that come specifically from science streaming needs,” said Buitrago. “So we are working with cosmology researchers, people doing image analysis for healthcare where they need to [handle] the high resolution images and also images in more than two dimensions and seeing how to address what are the best solutions for those specific use cases.”
The project partners are particularly enthused about harnessing AI for social good. Drug discovery, more accurate weather prediction, improved materials for increased solar energy generation and understanding large plant genomes to boost crop yields are just a few of the areas PSC expects will benefit from Neocortex as well as the upcoming Bridges-2 system (see slide below right for system details).
Both Neocortex and Bridges-2 — also built with HPE — will be deployed in the fall. “We’re launching two supercomputers in the same season,” Nystrom declared. “PSC has never done that before.”
As with Bridges-2, 90 percent of time on Neocortex will be allocated through XSEDE. “We’ll have a long early user period, but there’s also discretionary capacity for industry to work with us too, to use the world’s most advanced AI capability to develop their capacity for industrial competitiveness and for translational research,” said Nystrom.
There’s also a concerted focus, via the NSF-funded OpenCompass program, to collect and document best practices for running artificial intelligence at scale and communicate those to the open science community. This dovetails with a mission of PSC to support non-traditional users (from history, philosophy, etc.) and users who are just getting started with AI.
Neocortex will support the most popular deep learning frameworks and will be federated with PSC’s new Bridges-2 supercomputer, creating “a singularly powerful and flexible ecosystem for high performance AI, data analytics, modeling and simulation.”
Both Neocortex and Bridges-2 will be available at no cost for research and education, and at cost-recovery rates for industry users.
PSC will present a tutorial on AI hardware at PEARC (July 26-30) and will be talking more about the Neocortex system and what to expect. More details will be forthcoming at https://pearc.acm.org/pearc20/.