Intel embarks on a new adventure. Will it succeed? The chip manufacturer recently changed its direction, announcing it will develop a processor capable of fulfilling different tasks in parallel -- like Nvidia. Last May, Nvidia made history when it became the first chip company to cross the trillion-dollar market cap mark. Nvidia became the darling of Wall Street overnight mainly for one reason: the graphics processor that it developed in the 90s for gamers dovetailed perfectly with the generative artificial intelligence revolution. Nvidia fortified its position by building a virtual wall around itself, developing an operating system that worked only with its GPU, and producing a series of complementary products for major companies, such as supercomputers and cloud services.
Intel, Nvidia's veteran competitor, has also been trying for years to launch graphics processors for supercomputers and cloud services, as well as for artificial intelligence applications, but without much success. Its attempts to launch graphics processors suited to tasks that Nvidia GPUs handle successfully, such as training and maintenance of artificial intelligence models and supercomputing, met with considerable obstacles.
In the past year alone, Intel has cancelled one of its flagship development programs, for the GPU codenamed Rialto Bridge. Although the company launched the Gaudi2 processor last year -- developed in Israel for AI -- it has not so far been very successful for two reasons: the Israeli processor has not until recently proved better than its Nvidia counterpart, and its software environment was too new for most AI engineers.
Amazon was the main company to adopt Gaudi processors, but it is also simultaneously developing its own artificial intelligence chips that may, in the long run, challenge Intel's Gaudi.
Intel’s strategy has changed
And so, Intel made the laconic announcement that the Gaudi AI chip developed by Habana Labs, an Israeli company acquired by Intel, would reach end-of-life in two years. After an advanced version is launched soon, certain components will be integrated into another, future processor.
The disclosure of Intel's plans for its future graphics processor was accompanied by a statement from Jeff McVeigh, corporate vice president and general manager of Intel’s Super Compute Group, who announced that the company was canceling its original, ambitious plan to build a superchip combining graphics processing units (GPUs) with its central processing units (CPUs). Instead, Intel plans to develop a future graphics processor, codenamed Falcon Shores.
"My prior push, and emphasis around integrating CPU and GPU into an XPU was premature," McVeigh admitted at a press briefing held by Intel in May. " And the reason is, we feel like we are in a much more dynamic market than we thought even just a year ago." McVeigh explained that the generative artificial intelligence revolution and the increasing importance of large language models (LLMs) upended Intel’s plans.
In fact, the ten months that have passed since the launch of ChatGPT have brought about a fundamental change in the way Intel relates to the artificial intelligence revolution. Intel’s previous strategy, which posited that different chips should be launched for different tasks, such as for artificial intelligence, has been upended. Now, the company intends to develop one chip able to process all tasks in parallel, much like Nvidia's H100 chip. As part of this thinking, a second processor developed by Habana Labs, Greco, was dropped after it was realized that Gaudi performed operational tasks more efficiently.
Price of Nvidia chips an Intel advantage
Is this bad news for the Habana Labs development center in Caesarea and the Israeli company’s one thousand employees? Not necessarily. "Globes" has learned that Habana Labs, under COO Eitan Medina, is expected to lead the development architecture of the new Falcon Shores chip, and to integrate the communications components that enable chips to communicate with each other within data centers. Essentially, Habana Labs will become the equivalent of Mellanox (the Israeli company bought by Nvidia) within Intel, and will be responsible for all aspects of communications in supercomputers, data centers, and artificial intelligence.
In addition, most system-on-chips (SoCs) in the new GPU will come from an architecture developed originally for Gaudi by Habana Labs. However, the core of the new chip will come from another product developed at Intel: the veteran Xe technology, which was part of the projects that were shelved. Intel realized that there was no point in two or three departments developing graphics processors at the same time, and united them all under one roof.
The company is bolstered by the fact that the Gaudi2 chip, followed by Gaudi3, which will be launched in the first half of 2024, will take a chunk of Nvidia's large market share. Results published by tech analysis company MLCommons of its latest MLPerf Inference benchmark exercise revealed that Gaudi2 outperformed Nvidia's less advanced A100 chip when working with the GPT-J language model. However, Nvidia's more advanced chip, the H100, still scored higher. But performance isn’t everything in the industry. H100 chips are substantially more expensive. Therefore, given Gaudi's performance against the A100, Intel hopes to conquer the market of those willing to compromise on slightly less powerful chips.
In addition, although its activity has decreased over the years, Habana Labs is one of the only divisions at Intel still recruiting employees, and its workforce, which currently numbers about a thousand people, half of whom are in Israel, is actually expected to expand.
A checkered history of launches
If we consider Intel's history of developing graphics processors then, at least on paper, the chances of success are not in its favor. Even before cancelling Rialto Bridge, Intel had tried several other times to develop GPUs under different code names, such as Larrabee and Xeon Phi. It acquired AI software company Nervana Systems in 2016 in order to build its own graphics processor, then shut down its operations in favor of acquiring Habana Labs for $2 billion. At the same time, Intel hired Raja Koduri to head its GPU business unit and Uri Frank to lead core and client development (both no longer work for the company). However, production problems at Intel plants and low priority prevented Intel from improving its position during the last decade.
Another significant decision recently made by Intel was to have competitor TSMC manufacture its GPUs, enabling Intel to produce its next generation processors using a manufacturing process it had not yet been able to introduce itself: 5 nanometers. This move was made as part of Intel's transformation into a more open enterprise, willing to manufacture chips for the competition as well as allowing its competitors to make Intel chips.
It now remains to be seen how well Intel will succeed in its plans to launch its new graphics processor. The launch is planned for 2025, which is very late for a market in need of immediate solutions for training and operating AI models, and suffering the high prices charged by near monopoly Nvidia. However, Gaudi3 may surprise and challenge Nvidia on price as well.
Sandra Rivera, Executive Vice President and General Manager of the Data Center and AI Group at Intel, told "Globes", "Habana Labs continues to contribute significantly to engineering and to the execution of the roadmap of our artificial intelligence accelerator. Intel's decision to unite the roadmap of our accelerator and graphics processor into Falcon Shores highlights our confidence that Habana’s expertise and technological contribution are highly valued and are an inseparable part of our artificial intelligence strategy."
Published by Globes, Israel business news - en.globes.co.il - on September 20, 2023.
© Copyright of Globes Publisher Itonut (1983) Ltd., 2023.