Problems associated with GPU deployment – interminable order wait times, high prices and, particularly, dire need – are leading to new GPU access strategies.
An article in today’s Wall Street Journal, “Your Gaming PC Could Help Train AI Models,” reports that underused GPUs “inspire startups to stitch together virtual ‘distributed’ networks to compete with AI data centers.”
The article cites a number of company who are “among a burgeoning group of founders who say they believe success in AI lies in finding pockets of underused GPUs around the world and stitching them together in virtual ‘distributed’ networks over the internet,” stated the Journal. “These chips can be anywhere—in a university lab or a hedge fund’s office or a gaming PC in a teenager’s bedroom. If it works, the setup would allow AI developers to bypass the largest tech companies and compete against OpenAI or Google at far lower cost.”
This recalls the Folding@Home phenomenon (and similar efforts) that became widely used soon after the 2020 COVID-19 outbreak, in which scientists accessed idle distributed computing resources, starting with PCs and workstations that, in aggregate, delivered HPC-class compute for disease research.
One of the entrepreneurs cited in the article, Alex Cheema, co-founder of EXO Labs, stated that organizations around the world have tens and hundreds of GPUs that often are not being used – such as during non-business hours – that taken together have more GPU compute power than large AI data centers powered by hundreds of thousands of Nvidia GPUs.
The article notes that to date, virtual networks of GPUs have been scaled only to a few hundred chips, and that many technical and business barriers exist. Among them: network latency, data security, identifying contributors of idle GPUs, and the risk averseness of builders of costly AI models.
Still, sidestepping current high-cost GPU business models, be they on-premises, in a colo or in the cloud, will always catch the attention of IT planners.
The Journal quoted Paul Hainsworth, CEO of decentralized AI company Berkeley Compute, who said he is working a means of investing in GPUs as a financial asset that can be rented out. “I’m making a big bet that the big tech companies are wrong that all of the value will be accreted to a centralized place,” said Hainsworth, whose home page makes this offer: “Owners purchase GPUs that get installed and managed in professional datacenter(s), earning passive income through rental fees without needing any technical expertise.”
