An arms race over artificial intelligence and machine learning among chip makers continued to heat up on Tuesday, with IBM announcing its latest microprocessor aimed at the fast-growing market niche.
Among the highlights of IBM’s new Power9 line up of chips for servers are faster connections intended to help the processors work closely with graphics processing chips from Nvidia. IBM is including the second generation of Nvidia’s proprietary NVLink technology and a connection known as PCI-Express 4.0. Using the NVLink technology leads to a system that is almost 10 times faster at moving data than an Intel-based system, IBM claimed, though it provided no third party test results to back the boast.
“We can lift up the entire Library of Congress and run it around the chip in less than five seconds-tremendous bandwidth,” Bob Picciano, IBM senior vice president for cognitive systems, said.
IBM, Intel, Advanced Micro Devices, and other chipmakers are racing to offer increasingly speedy chips that can crunch the huge sets of data used in AI and machine learning such as image and voice recognition and credit card fraud prevention. Spending on AI-oriented microprocessors in servers will jump from $500 million last year to $10 billion in 2020, analysts at Bernstein Research have estimated. Competing with IBM and Nvidia, AMD (AMD) is pitching combinations of its newest Epyc server and Vega graphics chips for AI tasks, while Intel (INTC) wants to combine its Xeon server chips with products from two companies it recently acquired, Altera and Nervana Systems.
But most analysts agree that Nvidia’s graphics chips are the most popular among AI researchers, with appealing performance and a robust choice of software. IBM says it has an exclusive deal with Nvidia to run Nvidia’s NVLink technology. IBM (IBM) is also working with Google (GOOGL) and on building a new supercomputer called Summit including Power9 chips and Nvidia (NVDA) chips for the Department of Energy that could end up as the world’s fastest.
“That, according to the National Laboratories, is going to be the most powerful supercomputer in the world when it’s fully assembled and completed next year,” IBM’s Picciano said.
(Update: This story was updated on Dec. 6 to correct the capabilities of NVLink and PCI-Express 4.0)