Interest in artificial intelligence has exploded in recent months, ever since Microsoft and Google unveiled dueling chatbots. These rapid advances in AI don’t just happen, though. It takes unfathomable computing power to train and run these models. Intel hopes to get a bigger piece of this business with its upcoming Falcon Shores chip, and it has just revealed the first details of the design.
Intel was on hand at the ISC 2023 conference in Germany to talk about its high-performance computing plans. The company has confirmed that Falcon Shores will be ready for the next generation of large language models (LLMs), a category that includes OpenAI’s GPT, Google’s PaLM, and more. The chips will have 288GB of HMB3 memory and 8-bit floating-point processing. The popular Nvidia H100 Hopper card has “only” 80GB of RAM.
Nvidia is clearly in the lead regarding machine learning hardware, but AMD is expected to present a larger challenge later this year with the MI300. This will be AMD’s first design dedicated to high-performance computing, utilizing multiple chiplets on TSMC’s 5nm process, 3D die stacking, and on-package memory with ultrawide bandwidth. Intel, by comparison, has only just started focusing on developing hardware for ML applications.
Intel says it is preparing to wrap up work on Argonne National Lab’s Aurora supercomputer, which is based on the firm’s long-delayed Ponte Vecchio processor. Intel claims Ponte Vecchio is even faster than Nvidia’s H100, but no one except the researchers at Argonne National Lab would know. After that, it will be working hard to get Falcon Shores ready.
Intel’s current timeline calls for Falcon Shores to begin shipping in 2025. That gives Nvidia plenty more time to build up its lead, with AMD nipping at its heels. There may even be a new Nvidia ML processor by then. When Intel is ready to compete for machine learning data centers, it may find many big players have already settled on a computing partner.
Intel is banking on the AI obsession morphing into a long-term business opportunity. It’s not alone there—former crypto mining outfits that bought GPUs during the 2021-2022 metaverse boom are now trying to revamp operations to run ML workloads. There’s money to be made now, but there’s no way to know if we’ll have a new tech obsession in 2025.