The European Union plans to support its own AI startups by providing access to processing power for model training on the region’s supercomputers, announced and launched in September. . last monthAccording to the latest information from the EU, France’s Mistral AI is participating in an early pilot phase. But one early learning is that the program needs to include dedicated support to train AI startups on how to make the most of the EU’s high-performance computing. is.
“One of the things we’ve seen is that we don’t just provide access; facility — In particular, the skills, knowledge and experience we have at our hosting centers — to not only facilitate this access, but also to develop training algorithms that take full advantage of the architecture and computing power currently available at each supercomputing center. on how to develop and also in our machines,” an EU official said at a press conference today.
The plan is to establish a “center of excellence” to support the development of specialized AI algorithms that can run on EU supercomputers.
Rather than relying on the processing power provided by supercomputers as a training resource, AI startups may be accustomed to training their models using specialized computing hardware provided by US hyperscalers. It will be expensive. Access to high-performance computing for AI training programs is therefore being enhanced with support wrappers, said EU officials speaking in the background ahead of the formal ribbon-cutting. mare nostrum 5a pre-exascale supercomputer, goes live on Thursday at the Barcelona Supercomputing Center in Spain.
“We are developing a facility to help small and medium-sized enterprises understand how best to use supercomputers, how to access supercomputers, how to parallelize algorithms so that they can develop models in the case of AI. ” said a European Commission official. . “In 2024, we expect to see a lot more of this kind of approach than we do today.”
“AI is now considered a strategic priority for the Coalition,” they added. “Next to the AI Act, as AI becomes a strategic priority, we are providing innovation capabilities or enabling small businesses and startups to make the most of our machines and this public infrastructure. “We want to provide a major window of innovation.” ”
Another EU official confirmed that an “AI support center” was in the works, including a “special He said there would be a “prepared truck”. “What we need to realize is that the AI community hasn’t used supercomputers in the past decade,” they noted. “They’re not new users of GPUs, but they’re new to how to interact with supercomputers, so we need to help them.
“A lot of times the AI community comes from a huge amount of knowledge about how many GPUs you can put in a box. And they’ve been very good at it. What you have is a bunch of boxes with GPUs, and you need additional skillsets and extra help to scale out the supercomputer and exploit its full potential.”
The bloc has significantly increased its investment in supercomputers over the past five years, expanding its hardware to regionally located clusters of eight machines, interconnected via a Terabit network. We also plan to create federated supercomputing resources. Accessed in the cloud, it is available to users across Europe.
The EU’s first exascale supercomputers are also expected to come online in the next few years, with one in Germany (likely next year) and a second in France (expected in 2025). The European Commission also plans to invest in quantum computing, providing hybrid resources co-located with supercomputers and combining both types of hardware, so that quantum computers can act as “accelerators”. There are plans to acquire a quantum simulator that will As the committee states, it is a classic supercomputer.
Applications being developed on the EU’s high-performance computing hardware include projects that simulate Earth’s ecosystems to better model climate change and weather systems. destination earthand one more thing needs to be devised Digital twin of the human body This is expected to contribute to the advancement of medicine by supporting drug development and making personalized medicine possible. Leveraging his resources in supercomputing to launch his AI startup has recently been announced, especially after the EU president announced this fall that his AI model would have computing access to his training program. It is emerging as a strategic priority.
last monthThe bloc also announced what it called the “Large-Scale AI Grand Challenge.” This is a competition for European AI startups “with experience in large-scale AI models” and aims to select up to four promising domestic startups for a total of four. Access to millions of hours of supercomputing to support foundational model development. According to the European Commission, there will be a prize of 1 million euros to be distributed to the winners, who will be able to release their developed model or publish their research results under a non-commercial open source license. It is expected.
The EU already had a program that provided industry users with access to core hours of supercomputing resources through a project recruitment process. However, the coalition is increasing its focus on commercial AI with dedicated programs and resources, and there is an opportunity to incorporate its growing supercomputing network into a strategic power source for expanding ‘Made in Europe’ general purpose AI. They are intently aiming for this. Therefore, France’s Mistral (an AI startup that aims to compete with US infrastructure model giants like OpenAI and claims to offer “open assets” (if not completely open source)) will benefit early on. It seems no coincidence that it receives the Commission’s Supercomputer Access Program. (That said, the technology company, which just raised €385 million in Series A funding that includes US investors including Andreessen Horowitz, General Catalyst and Salesforce, is at the front of the line for computing giveaways.) That might raise some eyebrows, but hey, it’s another sign of the high-level strategic bets being made on “big AI.”)
The EU’s “Supercomputing for AI” program is still in its infancy, so it remains to be seen whether there will be enough benefits in model training to warrant reporting from dedicated access. (We reached out to Mistral for comment, but he did not respond as of press time.) But the committee’s at least hope is that by focusing support on AI startups, they will be able to move into high-performance computing. It is about being able to leverage investments. The construction of supercomputer hardware is increasingly being procured and configured with AI model training in mind, and this is due to the fact that local, hyperscalar-like US AI giants are starting at a disadvantage. This will be a competitive advantage for the AI ecosystem.
“We don’t have the massive hyperscalers that the Americans have when it comes to training this kind of basic model, so we’re using supercomputers and a new generation that’s becoming increasingly compliant with AI. “We intend to develop a supercomputer,” a committee official said. “The objective in 2024, not just with the supercomputers that we have now, is to move in this direction so that even more small and medium-sized businesses can use supercomputers to develop these basic models. It is to do.”
The plan includes acquiring “more dedicated AI supercomputing machines based on accelerators rather than standard CPUs,” they added.
Will the EU’s AI support strategy align with or diverge from certain Member States’ ambitions to develop national AI champions? We heard a lot about this during the recent difficult negotiations to develop the EU’s AI rulebook, in which France took the lead in pushing forward the AI rulebook.Regulatory carve-outs to the underlying model It drew criticism from small and medium-sized businesses. – As seen. But Mistral’s early presence in the EU’s supercomputing access program may suggest a consensus.