AI stands before the HPC industry as a beacon of great expectations, yet market research repeatedly shows that AI adoption is commonly stuck in the talking phase, on the near side of a difficult chasm to cross. In response, many vendors have built integrations of their server and storage platforms with AI hardware – usually Nvidia GPUs – designed to leapfrog the initial steps involved in AI implementations.
Penguin Computing this week announced its own market strategy for helping organizations jumpstart their AI journeys, launching a new AI practice that the company says will operate as a full-service consultancy, providing guidance on system design, building custom technology solutions, and delivering professional services and support.
In so doing, a senior strategist at Penguin Computing also looked ahead to a future in which GPUs may be augmented with special purpose inference processors as chips of choice for AI workloads (see below). He also discussed, in light of the widely shared observation that “AI is the new HPC workload,” the significantly different demands AI and traditional HPC workloads place on the processor.
About its new AI consultancy, Penguin Computing CTO Philip Pokorny told HPCwire, “We help customers build networks, rack layouts, and assist with figuring out the best deployment strategy, so they can focus on doing AI and not have to worry about the details.”
Since Penguin Computing was acquired by SMART Global Holdings in June, the newly formed subsidiary has been on a mission to double down on AI initiatives, said Pokorny, who leads the new AI consultancy. He added that the acquisition by SMART can provide financial resources that would help grow the Penguin-Computing-on-Demand service, a bare metal HPC cloud that is planned to be part of the menu of offerings provided by the AI practice.
Penguin Computing is known for its versatile product set that includes Linux servers, turnkey HPC clusters, as well as managed and hosted solutions. In addition to the aforementioned Penguin-Computing-on-Demand cloud service, the company also offers a managed service (on-prem at the customer site or collocated in Penguin Computing’s datacenters), available through its professional services arm.
“It’s that customizability of solutions and access to a broad base of technologies that we think will make our AI practice really valuable to customers who probably don’t have a lot of experience – or if they do have the experience, decide their time is better spent focusing on AI,” said Pokorny.
“We’re not announcing a point product and we’re not saying that we have a one-size-fits-all solution. What we’re saying is that if you want a solution customized to your needs, we have the expertise to build those customized solutions and we have the menu of capabilities to tailor that to each individual user’s needs.”
With its 20-year history as a turnkey Linux system builder, Penguin Computing’s objective is to set up its HPC and AI customers with the right infrastructure. The AI landscape is evolving fast and there are dozens of frameworks and libraries in the mix. Penguin Computing doesn’t want to tell customers what would work best for them or what framework they should optimize around. “We want to emphasize to customers they are free to choose whatever tool chain works for them, but we will have the experience in our AI practice to say, if you’ve chosen a given framework, then that restricts some choices you have in terms of operating system, fabric, and so on. We have the expertise to say, given your choices, this is what will work best,” said Pokorny.
Pokorny said Penguin Computing’s more popular hardware configurations for AI workloads include 4-GPU and 8-GPU servers, and it’s a certified reseller of the Nvidia DGX-1 system. He noted the company has done remarkably well deploying a large number of Nvidia DGX-1s and has seen its revenue attributable to Nvidia grow significantly over the past few years.
GPUs are currently a primary means of accelerating AI training and inference workloads, and Penguin Computing provides a number of GPU-centered options. But it is also exploring other architectures for AI, such as GraphCore and RISC-V-based silicon. Graphcore’s first card, based on its 16nm “Colossus” microarchitecture, is reportedly coming soon and Penguin Computing is in line to receive samples.
RISC-V is an open, free instruction set architecture based on established reduced instruction set computing principles. Penguin Computing bought one of the first stand-alone RISC-V Linux machines from SiFive earlier this year and has put some of them in the hands of its government customers, who are similarly interested in forward-looking CPU architectures. Pokorny sees RISC-V as “an open CPU architecture with the potential to dramatically lower costs and potentially give users access to specialized instructions for novel new architectures that could be enabled by not having a licensing tax.”
While these newer architectures don’t optimize for double-precision, the mainstay of HPC, that is not a mandate that the AI space shares. “When you look at the direction that tensor cores are going and the way that Nvidia is changing their architectures, double-precision is not where it’s at with respect to training, it’s all about smaller and smaller data sizes. It’s still very important in high-performance computing,” said Pokorny. He is open to the possibility, however, that the pattern-matching of AI could potentially replace or augment first principles physics-based simulations.
“There’s been a question posed but I don’t think it has been answered yet, in the HPC community,” said Pokorny, “which is that if you think of a baby, a baby learns about gravity by falling down and watching balls go up in the air and come down and roll down hill, and it doesn’t learn about gravity from force equals mass times acceleration and studying physics. So historically, we have tried to figure out what are the equations that underpin physics and then use those equations to build models that allow us to simulate physical phenomena. We do that starting from the ground up and it tends to be very expensive in floating-point cycles and fidelity. So the question that has been posed is, ‘Couldn’t we train a neural network to discover the rules of combustion or discover the rules of gravity by giving it a bunch of training sets and then asking it to tell us what would happen in a different scenario?’
“If it turns out that that actually works and is actually more efficient, you could imagine that could really alter the way we do HPC in the future. It opens an avenue to being much more energy efficient because it takes fewer computations to get the same fidelity of result. But it would also change to some extent whatever we call determinism. If you ran the same problem twice, you might get two different results due to some small fluctuations, so it’s an interesting problem to be stated and answered by the HPC community with regard to artificial intelligence.”
Pokorny also holds the view that while current GPUs are a dominant architecture for today’s most visible AI workloads, they may not continue that supremacy across the broader swath of AI coming out of the shadows. “If you take the example of say risk assessment – credit card risk and identifying fraudulent transactions – that was to some extent AI before AI was cool and as far as I know that work is done on traditional CPUs,” he told us. “You also have things like high-frequency trading which involve machines making decisions about recognizing patterns and then training on those patterns. To some extent that’s also AI before AI was cool.”
Pokorny is keen on the moves that Intel is making around FPGAs integrating into CPU sockets and points to Microsoft Azure’s dramatically speeding up its Bing search engine with FPGAs. “An FPGA has the promise of being a reconfigurable pile of hardware to do exactly what you need it to do very quickly,” Pokorny said. “It remains to be seen whether the cost is tractable, whether the performance pays off against the cost. But I also think that we’re seeing a global accessibility to ASICs to some extent strangely driven by the bitcoin miners. We’re seeing developments around making toolchains more approachable.”
Pokorny’s takeaway message: AI deserves the hype. “I don’t think it’s going to be as disillusioning as maybe we’ve seen in prior generations of artificial intelligence where we thought AI was just around the corner for the last 30 years. I think people are approaching this a lot more realistically. I think we have [done so] at Penguin Computing, and so the net of that is, I think commercial and Fortune 500 companies are going to be making more use of this to improve efficiencies and augment processes and do a better job of suggesting things. And it’s going to be really interesting to see in what ways AI transforms a lot of different workflows by being able to recognize patterns that maybe people weren’t able to recognize before.”
— Doug Black contributed to this report.