Nvidia has a new chip in the works for boosting artificial intelligence and other high-performance computing work: Grace, a design slated to arrive in mammoth supercomputers in 2023. Instead of accelerating conventional Intel-powered servers, though, the design includes its own built-in Arm processors.
Nvidia’s current brainiest chip, the A100, is typically yoked to Intel Xeon processors. Nvidia chips do the grunt work, but Intel chips oversee it. With Grace, named after pioneering programmer Grace Hopper, the company opted to embed several Arm Neoverse processor cores within the chip to speed up processing, said Paresh Kharya, an Nvidia senior director. The chip news arrived at Nvidia’s GTC 2021 conference this week.
The new chip should let AI customers run computing tasks that are vastly more complex than is possible with today’s chip designs, a step toward the general artificial intelligence that is the holy grail of today’s machine learning research, said Cambrian AI Researach analyst Karl Freund in a blog post.
The design illustrates Nvidia’s dramatic ascent — and Intel’s struggles. Even decades of dominance in technology don’t guarantee success when the rules of computing are constantly being rewritten. Your laptop likely comes with an Intel chip, but an Nvidia chip was more likely responsible for important AI work like filtering spam, improving image quality or recognizing your voice when you call your bank.
Not so many years ago, Nvidia was just a component supplier, a designer of graphics chips called GPUs to boost PC performance. Intel’s family of processors, or perhaps compatible rival AMD chips, shouldered most of the computing work. Intel, though, has struggled in recent years to keep pace with chip miniaturization and to capitalize on the exploding use of AI.
The result: Nvidia’s market capitalization vaulted over Intel’s, reaching $357 billion compared with Intel’s $278 billion. Much of the growth has been propelled by the fact that GPUs also turned out to be pretty good at AI work, specifically the computationally intense training process that builds the models that later run in data centers, PCs and phones.
Also in the ascendant is Arm, which licenses the chip designs and technology that power every smartphone, new M1-based Apple Macs and the world’s fastest supercomputer. Nvidia is seeking to acquire Arm for $40 billion, a move some rivals like Qualcomm object to. Grace’s integrated Arm chips let Nvidia read data from memory many times faster than with current designs, the company said.
Nvidia’s Selene machine, currently the world’s fifth-fastest supercomputer, pairs A100 chips with AMD Epyc CPUs. A 2023 Grace-based machine called Alps at Switzerland’s National Supercomputing Center should be seven times faster, Kharya said. The Los Alamos National Laboratory in the US also will buy a Grace-powered supercomputer.
Under new Chief Executive Pat Gelsinger, Intel is working to reclaim its manufacturing lead, planning to tap into others’ manufacturing abilities while it works on miniaturizing its circuitry inscribing technology.
Intel is building AI abilities into its main processors while working on dedicated hardware, too. It folded its Nervana chips operation, but its Habana AI acceleration processors are still under active development.
Atlan, Nvida’s new automotive chip
One hot area for AI chips is autonomous vehicles, whose self-driving algorithms rely on processing in camera imagery and other sensor data. It’s a core focus for Nvidia AI chip work, for example with its Orin chip scheduled to debut in 2022 vehicles.
Nvidia CEO Jensen Huang announced new processors for AI, graphics and supercomputing at the company’s GTC event.
Screenshot by Stephen Shankland/CNET
At GTC, Nvidia announced a new chip called Atlan with quadruple the performance. It should arrive in 2025 vehicles, said Danny Shapiro, Nvidia’s senior director of automotive work. Like Orin and Grace, Atlan relies on Arm cores, too.
Nvidia also announced a grander autonomous vehicle technology package called Hyperion 8. It combines two Orin processors with a host of sensors: eight exterior cameras, four exterior wider-angle fisheye cameras, three interior cameras, nine radar scanners and one lidar 3D scanner. The technology should arrive later in 2021.
Nvidia extended a partnership with Volvo, the companies said. Volvo plans to use Orin chips in its next-generation vehicles.
Intel has its own autonomous vehicle division, Mobileye. Tesla develops its own AI chips for its cars.
