Roaring demand for Generative AI enabling technology – mainly GPUs and GPU-accelerated systems – prompted Intersect360 Research to nudge up its 2023 HPC-AI forecast, which was first given at mid-year. The current spike will be relatively short, said Intersect360 CEO Addison Snell during his annual pre-SC23 market update last week. But for now, the generative AI gold rush is squeezing supplies and boosting prices.

“We definitely see a bump in 2023. We think that bump is continuing into 2024 and that results in additive $8.5 billion in spending coming into the market over the course of the five-year forecast, with a slight increase to the long-term growth rate. But really, most of that spending that’s additive in 2023 and 2024,” said Snell. “The market will not then fall down below where it was before. Everything is additive. Another effect is that cloud will continue its double-digit growth streak an extra year beyond what we had previously forecast through 2024.”

All netted out, Intersect360 is forecasting a 6.7% CAGR for HPC-AI through 2027, which will push the total market above $60 billion. Snell emphasized the numbers presented exclude spending by cloud providers for their internal HPC-AI resources.

“In the technology survey, we looked at data from the HPC and machine learning survey that we did earlier this year. We looked at the supplier trends, both for the system vendors and for supply chain, and what’s going on with GPUs. Dan and I concluded that these conditions have created a significant but temporary market bubble and this forecast is combining all of these factors into a revised five-year outlook, where the near-term gain is concentrated in spending on the servers because people are spending more on the server portion because GPUs are an expensive component,” said Snell, who delivered the market update jointly with Intersect360 chief research officer, Dan Olds.

The market update was delivered by a webinar, and Intersect360 has made the slide deck and the video recording of the roughly 90-minute presentation available from its website (registration may be required). Many of the trends are familiar – spending growth on HPC-AI in the cloud will remain strong as will spending on storage, reported Intersect360.

Forecasting the post-pandemic HPC-AI landscape remains a difficult task. For starters, HPC and AI are now inextricably mingled, and the blurred the edges between what were once distinct markets can be tough to discern. Likewise, macro global influences – notably war (Ukraine, Gaza), ongoing geopolitical tension, shifting trade restrictions – further complicate things.

Shown below (slide) is Intersect360’s broad definition of the HPC-AI market

In addition to presenting the overall market forecast, Snell and Olds dug into some of top line results from Intersect360’s 2023 end-user study conducted in the Q2 of this year.

The brief portraits of server, storage, and interconnect markets were mostly familiar (see HPCwire coverage of Intersect360 mid-year report). Intersect360 didn’t present, strictly-speaking, market share numbers for the server community but the top-named server suppliers among users to its survey. Dell was the top performer, followed by HPE, Supermicro, and Nvidia (slide below).

Snell said HPE likely still has the largest market share (total spend, $) because it sells so many larger systems, while Dell has also extensive market penetration but sells a larger number of smaller units. He contended that Nvidia, often thought of as only component supplier (GPU, Interconnect (Mellanox) etc.), should now also be regarded a server/full system provider, not only because of its DGX line of systems, but also because of wide-spread use of systems based on its reference designs.

CPU use is another interesting area. X86 architecture still dominates the market but AMD is cutting into Intel’s dominance (slide below). Interestingly RISC-V remains an outlier in terms of expected future use by survey respondents while ARM is growing but modestly. Intersect360 even suggested Intel’s market share could dip below 50% in the future.

A majority of HPC-AI systems are now shipped with accelerators, usually GPUs. Nvidia clearly remains the dominant player but attitudes are changing as the GPU supply squeeze persists, and more AMD (M1300) and Intel GPUs (Ponte Vecchio) start shipping.

Olds noted, “There are CEOs out there who are ringing up their CIOs and everybody down the chain saying, ‘We’ve got to be on this generative AI now’ and they salute and say ‘Yes, absolutely’ and are going out and casting around for any hardware so they can report back and say, ‘Yes, we’re on it.’ However, we are tapped out as an industry in terms of supply for Nvidia and AMD GPUs, that is driving the price is higher, and it’s going to keep them high.”

The flip side, said Olds, that’s going to create opportunity for substitutes from current and new players. Fox example, AMD’s Instinct M1300, just coming to market, is gaining wins. Snell and Olds both expressed surprise over Intel’s Gaudi2 accelerator’s lack of showing in the Intersect 360 survey and wondered if supply was an issue. In any case, the current GPU squeeze is likely to float many boats. (See HPCwire article, The Great GPU Squeeze is Upon Us)

Both Snell and Olds suggest market accelerator demand will ease as user organizations move from early fascination with generative AI to seeking concrete business cases. They pointed out that even now, many HPC-AI systems are either under-utilized overall or their accelerator portions are under-utilized. An exception is in government sectors whose HPC-AI systems tend be more fully-used; this is certainly true at major supercomputer centers.

Snell and Olds suggest the current seller’s market will rebalance and that the increases in supply, will drive prices down. The tough question is when.

Intersect360 believes the migration of HPC-AI workloads to the cloud remains a strong trend but with caveats. Currently 55% of the HPC-AI survey respondents use the cloud “at least” sometimes and that’s going to jump to 77% in five years. The leading HPC-AI cloud providers are familiar. Amazon Web Service dominates (53%) with Microsoft Azure (24%) and Google Cloud Platform (15%) in the top three spots in the Intersect360 study. Next were Oracle (2%), IBM (1%), and others (4%)

“AWS continues as number one, Azure continues as number two, and GCP continues is number three. The rest continue to not move that much. There’s this has been pretty stable. If anything, I would say AWS and Azure had been gaining relative to everything else,” said Snell.

HPC-AI Cloud use, said Snell, is strongly skewed to smaller, commercial users, but that’s likely to change as the public sector is poised to spend more in the cloud. Currently, HPC-AI users (survey respondents) report using the cloud for 12% of their HPC-AI workloads and expect that to increase to 22% in five years.

The big caveat, contends Intersect360, is that running HPC-AI workloads in the cloud remains expensive.

Olds said, “I believe we’re going to see an inflection point where cloud usage isn’t going to penetrate much more of HPC AI, and it’s certainly not going to take over all of it. There’s just economic reasons for that. It’s too expensive. You lose some control. There’s a litany of reasons behind that. But Cloud is great for overflow for testing, for development, for getting at scale testing when you need it. For things like that.”

There was much more discussed during update, such as operating system use and likely changes (think Rocky Linux), storage trends, container use proliferation, liquid cooling adoption, composability, and more. It’s best to watch the presentation directly.

Link to presentation and slides, https://www.intersect360.com/presentation/intersect360-research-pre-sc23-hpc-ai-market-update/