Snowcap Compute raises $23 million for superconducting AI chips
作者：Stephen Nellis
Snowcap Compute Chief Technology Officer Quentin Herr, Chief Science Officer Anna Herr, and CEO Mike Lafferty pose for a picture at the office of venture capital firm Playground Global in Palo Alto, California, U.S., in this undated handout photo. Courtesy Snowcap Compute/Stephanie Cowan/Handout via REUTERS Purchase Licensing Rights, opens new tab
- Snowcap aims to outperform AI systems with superconducting chips
- Superconducting chips face cooling challenges but promise efficiency
- Former Intel CEO joins Snowcap board
SAN FRANCISCO, June 23 (Reuters) - Snowcap Compute, a startup working on building artificial intelligence computing chips using superconducting technology, on Monday raised $23 million and said that the former CEO of Intel will join its board.
Snowcap aims to build computers that could one day beat today's best artificial intelligence systems, while using a fraction of the electricity. To do that, Snowcap plans to use a new kind of chip made with superconductors, which are materials that allow current to flow without electrical resistance.
Sign up here.
Scientists understand superconductors well and have theorized about making computer chips with them since at least the 1990s, but have faced a major challenge: To work, the chips need to be kept very cold in cryogenic coolers which themselves consume a lot of electricity.
For decades that made superconductor chips a nonstarter, until AI chatbots ignited huge demand for computing power at the same time that conventional chips are hitting the limits of how much performance they can wring from every watt of power and are taxing electricity grids.
Nvidia's forthcoming "Rubin Ultra" AI data center server due in 2027, for example, is expected to consume about 600 kilowatts of power. That means operating that single server at full capacity for one hour would consume about two thirds the average power that a U.S. house uses in a month.
In that kind of changed world, dedicating a portion of a data center's power needs to cryogenic coolers makes sense if the performance gains are good enough, said Michael Lafferty, Snowcap's CEO, who formerly oversaw work on futuristic chips at Cadence Design Systems
(CDNS.O), opens new tab. Snowcap believes that even after accounting for energy used in cooling, its chips will be about 25 times better than today's best chips in terms of performance per watt.
"Power (efficiency) is nice, but performance sells," Lafferty said. "So we're pushing the performance level way up and pulling the power down at the same time."
Snowcap's founding team includes two scientists - Anna Herr and Quentin Herr - who have done extensive work on superconducting chips at chip industry research firm Imec and defense firm Northrop Grumman
(NOC.N), opens new tab, as well as former chip executives from Nvidia
(NVDA.O), opens new taband Alphabet's
(GOOGL.O), opens new tabGoogle.
While the chips can be made in a standard factory, they will require an exotic metal called niobium titanium nitride that Lafferty said depends on Brazil and Canada for key ingredients. Snowcap plans its first basic chip by the end of 2026, but full systems will not come until later.
Despite the long development timeline, Pat Gelsinger, Intel's
(INTC.O), opens new tabformer CEO who led the investment for venture firm Playground Global and is joining Snowcap's board, said the computing industry needs a sharp break from its current trajectory of consuming ever more electricity.
"A lot of data centers today are just being limited by power availability," Gelsinger said.
Also joining the funding round were Cambium Capital and Vsquared Ventures.
Reporting by Stephen Nellis in San Francisco; Editing by Christopher Cushing
Our Standards: The Thomson Reuters Trust Principles., opens new tab
关于《Snowcap Compute raises $23 million for superconducting AI chips》的评论
暂无评论