Intel unveils AI chips to compete with Nvidia and AMD
Sign up now: Get ST's newsletters delivered to your inbox
Intel’s new product with the most to prove may be the Gaudi 3, the latest instalment of a line that competes with Nvidia’s industry-leading H100.
PHOTO: BLOOMBERG
Follow topic:
San Francisco – Intel, the biggest maker of PC processors, have announced new chips for personal computers and data centres that the company hopes will give it a bigger slice of the booming market for artificial intelligence (AI) hardware.
The line-up includes updated Xeon server chips – the second overhaul of that processor in less than a year – that use less electricity while boosting performance and memory, the company said on Dec 14. Intel’s Ultra Core chips for laptops and desktop computers, meanwhile, will let PCs process AI functions directly.
Intel’s new product with the most to prove may be the Gaudi 3, the latest instalment of a line that competes with Nvidia’s industry-leading H100. These chips, known as AI accelerators, help companies develop chatbots and other rapidly proliferating services. Gaudi 3 is on schedule for release in 2024, and Intel said that it will outperform the H100.
Intel shares rallied as much as 5.6 per cent in New York trading following the announcement, outpacing a 3.1 per cent gain by the Philadelphia Stock Exchange Semiconductor Index.
Intel chief executive Pat Gelsinger is counting on AI features to help reinvigorate growth at the company, which has suffered from past missteps and a broader PC slump. But he faces tougher competition than ever. Long-time rival Advanced Micro Devices (AMD) has snatched market share in PCs and servers, and some of Intel’s largest customers are now designing their own chips in-house.
At the same time, Nvidia has become a dominant force in data centre chips
AMD is also playing catch-up with Nvidia in AI accelerators. Its version of that product, the MI300, will debut in 2024. AMD unveiled the chip at an event last week and announced that the market for AI accelerators could climb to more than US$400 billion in the next four years.
Intel looks to gain an edge by shifting more AI processing towards devices, rather than data centres. The new Ultra Core chips, which do just that, will be available in roughly 230 PC models from manufacturers such as Dell Technologies and Samsung Electronics, starting as soon as Dec 14, Intel said.
Another shift could help Intel. For now, the AI industry has concentrated on developing chatbots and services through a process called training, which involves bombarding the software with data. In the future, companies may be more focused on actually running their completed software – something that can be handled with Xeon processors in data centres and with PC chips, Intel said.
“A few people create models – lots of people use them,” Mr Gelsinger told the audience at an event in New York.
AI will be handled by PCs and other gadgets because of the high costs of data centres, he said, adding that sending data over the Internet also limits the responsiveness of systems.
But running AI software on a laptop can quickly drain the battery. To mitigate that risk, the new Core parts will run more efficiently and – in their highest-end configuration – provide more than 10 hours of life even during the most demanding tasks, Intel said.
Intel’s Xeon range will get new components that make them 42 per cent better at running AI-related workloads than the previous generation, the company said. That predecessor was rolled out in January 2023.
Overall, the component will be 36 per cent better when measured by performance per watt of electricity, Intel said, adding that even more new versions of the Xeon line will be ready in the first half of 2024, part of its push to speed up product releases.
Mr Gelsinger aims to turn around the company and restore it to prominence in the semiconductor industry, which he believes will grow to a trillion dollars in sales by the end of the decade. He assured the audience on Dec 14 that his comeback plan is working.
“We’re going to do a decade of semiconductor work in four years,” he said. “That’s pretty crazy. We’re on track, baby.” BLOOMBERG

