How Intel’s new AI Gaudi 3 chip compares with Nvidia chips

Trying to unseat Nvidia

Intel’s moves come as the vendor and others chipmakers including AMD keep trying to compete with Nvidia in the face of Nvidia’s dominant position in the AI hardware market.

Most recently, Intel banded with other chip vendors, including Arm and Qualcomm, to create the Unified Acceleration Foundation (UXL).

UXL aims to decrease the dependence on GPUs and provide options for enterprises looking to power AI processing.

While Nvidia’s dominance comes from having the “first mover” advantage due to its widely used CUDA framework, many enterprises using CUDA face a significant challenge, said Ben Carbonneau, an analyst at Technology Business Research. CUDA is a parallel computing platform and programming model created by Nvidia.

“The proprietary nature of CUDA poses vendor lock-in risks, which can inhibit scalability,” Carbonneau said.

In response to the ubiquity of CUDA, AMD — a competitor to both Nvidia and Intel — has invested in its own ROCm platform and developer ecosystem — an open source stack for GPU computing — while providing CUDA porting capabilities that give developers the option to migrate CUDA code so that it can run on ROCm.

“This open, multi-vendor approach helps eliminate vendor lock-in,” Carbonneau said.

Gaudi 3 vs. H100 vs. Blackwell

The Gaudi 3 AI accelerator also helps enterprises avoid becoming tied to a single vendor, Intel said.

The chip can power AI systems with up to tens of thousands of accelerators connected through standard ethernet, the vendor said.

It provides open, community-based software and industry-standard ethernet networking that enables enterprises to scale from a single node to clusters, super-clusters, and mega-clusters, according to Intel.

It will be available to OEMs such as Dell Technologies, Lenovo and Hewlett Packard Enterprise by June, Intel said.

While Gaudi 3 can compare to Nvidia’s H100 in terms of speed and power, it isn’t fully comparable with Nvidia’s new Blackwell GPU, according to industry analysts.

“They’re not 100% comparable,” said Alvin Nguyen, an analyst at Forrester Research. For instance, Nvidia’s Blackwell starts at 700 watts and Intel’s Gaudi 3 starts at 600 watts.

They’re saying it’s not just about pure performance. You have to look at energy, you have to look at sustainability.

With the new chip, Intel is focusing on not only performance, but also other aspects of the data center, such as reducing energy consumption, Nguyen said.

“They’re saying it’s not just about pure performance. You have to look at energy, you have to look at sustainability,” he added.

Intel is also emphasizing the idea of AI everywhere by including AI in all of its products, Nguyen added.

Gaudi represents Intel’s highest level of AI acceleration that can be found not only inside the chip but also in the vendor’s Xeon Server processors, PCs and Core Ultra processor series.

Challenges with Gaudi and edge offerings

One challenge with Gaudi 3 is that Intel failed to show significant large-scale wins, said Ryan Shrout, president and GM of tech performance testing firm Signal65 and a former Intel chief performance strategist.

“We don’t have huge customers coming up on stage saying, ‘We’re investing half a billion dollars or $2 billion,’ or, “We’re going to buy 50,000 of these chips.’ And that’s really what I think Intel needs to see, Gaudi as a real competitor in the GPU space versus Nvidia or AMD,” he said.

Moreover, while performance isn’t everything, some enterprises might be willing to pay more for performance if it makes a difference in end-user applications, Nguyen said.

“It’s about making the clients happy and serving what they need from a business perspective,” he said. “People are going to be willing to pay a little bit more to have a better experience for, ultimately, end users.”

Other than Gaudi 3, Intel updated Xeon 6 Processors that offer performance-efficient tools to run current GenAI technologies like retrieval-augmented generation. The processors are aimed at use in data centers, the cloud and edge computing.

Intel also said its next-generation Intel Core Ultra client processor family, launching later this year, is geared toward the next generation of AI PCs, an AI edge application.

With its edge releases, Intel is looking to provide enterprises with connectivity, Nguyen said.

AI systems need a lot of data, and being able to transfer that data is essential for Intel’s AI at the edge strategy, he said.

“They’ve got the ability to get data back and forth to devices at the edge, where it’s closer to users,” Nguyen said.

Also at Intel Vision, Intel and Anyscale said that the latest version of Anyscale’s Ray platform, Ray 2.10, adds support for Gaudi 3.

Esther Ajao is a TechTarget Editorial news writer and podcast host covering artificial intelligence software and systems.