AMD Chairman and CEO Lisa Su speaks on the 2023 Consumer Electronics Show.
AFP via Getty Images
Europe’s strongest supercomputer is situated in a former paper factory in a small town in Finland, just 320 kilometers south of the Arctic Circle. Most data centers today are equipped with hundreds of chips from Nvidia, the chip of selection for developing AI applications. But Lumi, the $160 million computer named after the Finnish word for snow, comprises 12,000 MI250X graphics processors from rival chipmaker AMD.
It was these chips that sparked a collaboration between AMD and Silo AI of Helsinki, Finland, a research startup with 300 PhD students and researchers that develops AI software for big firms that may ultimately be utilized in a Acquisition valued at $665 million last week – the most important acquisition of an AI startup in Europe since Google bought Deepmind 400 million US dollars in 2014.
Silo AI, which received a grant to make use of AMD’s Lumi supercomputer and chips, used that computing power for a side project: training an open-source large language model (LLM) based on Finnish and English, in addition to one other model for Nordic languages. Using a supercomputer for this was unusual because GPU-dedicated clusters are higher at handling the maths used for AI, but Peter Sarlin, CEO and co-founder of Silo AI, said Forbes that he realized that AMD’s chips were as much as the duty. Not only that, they were cheaper too.
“We have shown that training can be scaled to AMD,” said Sarlin Forbes“We had a big team and invested a lot of effort, but given that, we saw that AMD was very competitive.”
Vamsi Boppana, senior vice chairman of AMD’s artificial intelligence group, said Forbes that Silo AI’s experience developing software for firms comparable to Rolls-Royce, Unilever and Europe’s largest insurer Allianz matches with AMD’s work. “Customers come to us and ask, ‘What’s the best way to run this model on AMD?’ And we can help them do that,” says Boppana, adding that Silo AI could also help shape future AMD GPU designs.
“A big incentive was the quality of the team, their cutting-edge work… another reason was the work they did on AMD platforms,” Boppana said.
The purchase of Silo AI could help the corporate close the gap with Nvidia, its biggest competitor – especially in software. AMD’s silicon is inside Touch distance of Nvidia’s older chips, as does its corporate headquarters in Santa Clara, Calif., however the chip giant has an enormous head start in software, having spent nearly twenty years developing Cuda, a platform utilized by thousands and thousands of developers to place together AI tools and apps. “AI is not just about hardware, it’s also about software and its ever-increasing complexity,” said Patrick Moorhead, principal analyst at Moor Insight & Strategy.
In addition to success in hardware, this lead in software has helped Nvidia’s market capitalization explode to $3 trillion for the reason that launch of ChatGPT in November 2022. AMD’s share price itself has risen a more modest 134% over the identical period, giving the corporate a worth of $286 billion.
AMD’s billionaire CEO Lisa Su has already tried to make it easier for developers to make use of the corporate’s AI chips, spending $125 million last 12 months on plenty of smaller AI labs, including AI tooling startups comparable to Nod AI and Mipsology.
AMD has also partnered with a few of Silicon Valley’s biggest firms to support Triton, an OpenAI-powered alternative to Cuda. Thanks to those efforts, AMD’s GPU revenue forecast has surpassed $4 billion this 12 months, doubling its 2015 forecast. last 12 months. Nvidia’s high prices and provide chain constraints could also encourage more firms to modify if AMD could make its chips more user-friendly.
Sarlin’s team is one in all a handful of mostly small AI research projects and startups that depend on AMD hardware. However, AI research center Hugging Face recently partnered with AMD to make it easier to run AI models on its chips, while Meta and OpenAI are also signed orders for AMD’s latest chip, which is meant to compete with Nvidia’s flagship H100. But the competition also goes beyond Nvidia: Microsoft, Amazon and Alphabet are also constructing their very own AI chips, and Google’s Gemini chatbot was trained on its tensor processing units.
Silo AI’s consulting work for European firms brought in $22 million in revenue last 12 months and helped the corporate, co-founded in 2017 by Sarlin and five other entrepreneurs and researchers, change into one in all Europe’s largest AI research labs. The company received a $17 million funding round in 2022 at an undisclosed valuation from Swedish private equity investor Altor.
Boppana says the Silo team’s technical skills and experience applying technology in enterprises contributed to the deal’s success. “It’s one thing to build this cool model, but quite another to deploy it and make it work,” he says.
