Ampere leads a pack of AI startups to challenge Nvidia's 'unsustainable' GPUs

AI Platform Alliance probably has Jensen Huang in tears...of laughter

History is full of examples of smaller players coming together to take on larger competitors. With the launch of the AI Platform Alliance this week, Arm CPU vendor Ampere aims to do just that in a bid to challenge Nvidia's hegemony.

The alliance brings together several familiar chip startups, including Cerebras Systems, Graphcore, Furiosa, Kalray, Kinara, Luminous, Neuchips, Rebellions, and Sapeon with the goal of promoting an open AI ecosystem.

"The key problem we see today is that incumbent solutions don't provide these things. It is currently a closed ecosystem on the hardware and software side, it is not efficient due to the large amounts of power being consumed, and it's costly," Ampere chief product officer Jeff Wittich told The Register in an email. "This makes it hard for people to scale AI to meet their needs without it being costly and unsustainable."

No one chipmaker can address this problem on their own, he argued. "We wanted to band together to create full platform solutions that are implementable." So it's no surprise that the alliance is focused so heavily on building out an ecosystem of AI systems more efficient than GPUs.

The explicit reference to "GPUs" in the release and Wittich's remarks regarding closed software and hardware ecosystems is almost certainly a shot at Nvidia. The chipmaker's latest H100 GPUs are rated for 700 watts and, loaded into their HGX mainboards, can consume upwards of 10kW at the system level.

Working together

There are a number of ways chipmakers are attempting to curb the power requirements of AI. Efficiencies of this sort can be made at hardware, software, or standards levels.

For instance, the chipmakers could come together to establish common form factors so that original design manufacturers can develop a chassis capable of supporting multiple vendors' accelerators. On the software front, the chipmakers could align around a common set of code tools and frameworks so that models run regardless of the underlying hardware architecture.

The alliance's first order of business fix the hardware issue.

"Our initial focus is to create a wide variety of hardware solutions that can utilize existing OEM and ODM platforms, so they can be readily deployed today," Wittich said. "It's about creating choice. Today, customers don't have a lot of choices, and people tend to default to the one solution they know. But, there are better solutions out there that are more efficient and cost effective, and our goal is to make those solutions available to end customers, so they can adopt them."

However, Cerebras CEO Andrew Feldman tells The Register that the development and release of open foundation models will be a key enabler as well. We also reached out to Graphcore to get their take on the alliance, but the chipmaker declined to comment.

Ampere expects the Alliance to attract additional members in the coming weeks, with any AI company building hardware welcome to join the party. However, it remains to be seen whether wider availability of AI systems from alliance members will be enough to chip away at Nvidia's stranglehold on the market.

The elephants in the room

While experts disagree on just how large Nvidia's claim over the AI space is — estimates range from anywhere from 75 to 90 percent share — there's no question the company leads the market. And as we discussed last week, recent developments around the company's product roadmap are likely to extend this lead, so long as they can make the networking work.

While the AI Platform Alliance includes many familiar AI startups, two of the largest companies working to steal share away from Nvidia are missing from the list: Intel and AMD.

Both companies are heavily invested in the development of their own AI hardware and software. What's more, they have both championed open source frameworks and tools, including many that work with third-party platforms.

At Intel Innovation in September, Intel CTO Greg Lavender talked up the chip giant's AI contributions, including its open source OneAPI, OpenVINO, and SYCL runtimes. SYCL is particularly interesting as it isn't limited to Intel's hardware and instead aims to be an open alternative to Nvidia's CUDA.

AMD also leans heavily on the open source community when it comes to software. The company recently teamed up with Hugging Face to expand the catalog of open source models available to run on its Instinct and Alveo accelerators.

Meanwhile, on the hardware front, both chipmakers have extensive experience — not to mention resources — working with ODMs to standardize form factors for a variety of systems.

Asked whether the alliance would be open to adding either company to its ranks, Wittich said "anyone is welcome to apply," and that the initial crop of partners was pulled from vendors Ampere had worked with previously. ®

More about


Send us news

Other stories you might like