EU to expand support for AI startups to tap its supercomputers for model training

EU to expand support for AI startups to tap its supercomputers for model training

EU to expand support for AI startups to tap its supercomputers for model training

Natasha Lomas@riptari / 1:07 AM GMT+7•December 20, 2023

 Comment

Image Credits: Vladimir_Timofeev / Getty Images

A European Union plan to support homegrown AI startups by providing them with access to processing power for model training on the bloc’s supercomputers, which was announced back in September and kicked off last month, has seen France’s Mistral AI participate in an early pilot phase, according to an update from the EU. But one early learning is the program needs to bake in dedicated support for AI startups to train them on how to get the most out of the EU’s high performance computing.

“One of the things that we have seen is the need, not only to provide access but, to provide facility — especially skills, knowledge and experience that we have in the hosting centres — on how this access can be not only facilitated but to develop training algorithms that are using the best of the architecture and the computing power that is available right now in each supercomputing center and in our machines,” said an EU official speaking during a press briefing today.

The plan is for “centers of excellence” to be set up to support the development of dedicated AI algorithms that can run on the EU’s supercomputers, they added.

AI startups are more likely to be accustomed to using dedicated compute hardware provided by U.S. hyperscalers to train their models than tapping the processing power offered by supercomputers as a training resource. So the high-performance computing access for AI training program is being augmented with a support wrapper, according to EU officials who were speaking on background ahead of the official ribbon cutting for MareNostrum 5, a pre-exascale supercomputer that will be inaugurated Thursday at the Barcelona Supercomputing Center in Spain.

“We’re developing facilities for our SMEs to be able to understand how to best use the supercomputers and how to access the supercomputers and how to parallelize their algorithms in the case of AI to be able to develop their models,” said a Commission official. “As of 2024, we expect much more of these kinds of approaches than we have right now.”

“AI is considered now a strategic priority for the Union,” they added. “With AI becoming a strategic priority, next to the AI Act, we’re providing the innovation capacity — or we want to provide a large innovation window for our SMEs and startups to be able to best use our machines and this public infrastructure we’ve been creating so that they can compete internationally in developing safe, trustworthy and ethical AI algorithms.”

An “AI support center” is on the way, confirmed another EU official — saying this will have “a special track” for SMEs and startups to get help to get the most out of the EU’s supercomputing resource. “What we need to recognise is that the AI community have not been using supercomputers for the last decade,” they noted. “They’re not new users to GPUs but they’re new users to how to engage with a supercomputer and therefore we need to help them.

“In many cases the AI community comes out of a huge knowledge about how many GPUs can you get into one box? And they’ve been very good at that. But what we have on the supercomputers is a lot of boxes with GPUs and there’s some extra skill-sets and some extra help that’s needed in order to scale out and use the supercomputer to its full potential.”

The bloc has been substantially stepping up its investment in supercomputers over the past five+ years — growing the hardware to a cluster of eight machines located around the region, which it also plans to interconnect, via terabit networks, to create a federated supercomputing resource that will be accessible in the cloud so it’s available for users all over Europe.

The EU’s first exascale supercomputers are also due to come on stream in the next couple of years — one in Germany (likely next year) and a second machine in France (expected in 2025). The Commission also has quantum computing investment on the slate, with plans to acquire a clutch of quantum simulators that will be co-located with supercomputers to provide a hybrid resource that combines both types of hardware so the quantum computers can act as “accelerators” for the classical supercomputers, as the Commission tells it.

Applications that are being developed atop the EU’s high-performance computing hardware include a project to simulate Earth’s ecosystems to better model climate change and weather systems, called Destination Earth, and another to devise a digital twin of the human body — which it’s hoped will help further medical science by supporting drug development and even enabling personalized medicine. Using its supercomputing resources to fire up AI startups specifically has emerged as a more recent strategic priority after the EU president’s announcement of the compute access for AI model training program this fall.

Last month, the bloc also announced what it branded a “Large AI grand challenge”: A competition geared toward European AI startups “with experience in large-scale AI models” that aims to select up to four promising homegrown startups that will get a total of 4 million hours of supercomputing access to support development of foundational models. A €1 million prize pot is also earmarked for distributing to the winners — which are expected to release their developed models under an open source license for noncommercial use, or through publishing their research findings, per the Commission.

The EU already had a program to provide industry users with access to core hours of supercomputing resource, via a call for projects process. But the bloc is amping up attention on commercial AI with dedicated programs and resourcing — spying the opportunity to gear its growing supercomputing network into a strategic power source for scaling “Made in Europe” general-purpose AI. It thus looks to be no accident that France’s Mistral — an AI startup that’s aiming to compete with U.S. foundation model giants like OpenAI, and which claims to offer “open assets” (if not to be fully open source itself) — is an early beneficiary of the Commission’s supercomputer access program. (Albeit, it may raise a few eyebrows that a tech firm that just pulled in €385 million in Series A funding, including from U.S. investors such as Andreessen Horowitz, General Catalyst and Salesforce, is at the front of the queue for a compute freebie in the EU. But, well, it’s another sign of the high-level strategic bets being made on “large AI.”)

It’s still early days for the EU’s “supercompute for AI” program, so it’s unclear whether there’s much model training upside to report off of dedicated access as yet.

Reached for comment, Mistral co-founder and CEO Arthur Mensch told us: “We used Leonardo [one of the EU’s current gen supercomputers, which is located in Bologna, Italy] to run a few small experiments this summer as the cluster was ramping up. It was a good collaboration in which we gave a lot of feedback and could get some interesting results. All our models were trained on our own cluster though.”

Clearly, though, the Commission’s hope is that by funnelling support at AI startups to tap into its investment in high-performance computing, combined with building out supercomputer hardware that it says will, increasingly, be procured and configured with AI model training in mind, this will translate into a competitive advantage for a local AI ecosystem that is starting on the back foot versus hyperscaler-proximate U.S. AI giants.

“Since we do not have the large hyperscalers that Americans have in the case of training these kinds of foundational models we use our supercomputers and we will develop a new generation of supercomputers that will be more and more AI compliant,” noted a Commission official. “Not only the ones that we have right now but, as of 2024, the purpose would be that we will go in this direction — and have even more of our SMEs use the supercomputers for developing these foundational models.”

The game plan will include acquiring “more dedicated AI supercomputing machines, that will be more based on accelerators rather than the standard CPUs,” they added. The newest machine — Barcelona-based MareNostrum 5 — has been designed to be used by AI developers, per the Commission.

Whether the EU’s AI support strategy dovetails with or diverges from the ambition of certain member states to foster national AI champions — something we heard a lot about about during the recent fraught talks to set the bloc’s AI rulebook, in which France led a push for a regulatory carve-out for foundational models that garnered criticism from SMEs — remains to be seen. But the early presence of Mistral in the EU’s supercomputing access program may suggest an alignment in the thinking.

This report was updated with comment from Mistral

EU to let ‘responsible’ AI startups train models on its supercomputers

 

France’s Mistral dials up call for EU AI Act to fix rules for apps, not model makers