Intel merging Gaudi tech into GPU product lines

Interesting-- sounds like Gaudi’s reasonably interesting if your problems don’t hit the highest scales like LLM training does… Give them three or four more years, they may have something.

The shortage of Nvidia’s GPUs has customers searching for scrap heap to kickstart makeshift AI projects, and Intel is benefitting from it. Customers seeking quick implementation of AI projects are looking at alternatives to GPUs, said David Zinsner, Intel’s chief financial officer, during a meeting with financial analysts at the Citi Global Technology Conference this week.

“The challenges in getting GPUs — I think we see more customers taking a look at Gaudi as an alternative. And in addition, the price points are better and more attractive,” Zinsner said.

Zinsner also remarked that Gaudi and Gaudi2 chips are shipping, with Gaudi3 coming out in a “reasonable period of time.”

During an earnings call last month, CEO Pat Gelsinger said the company has a $1 billion pipeline for AI chips. Zinsner said the pipeline is unrelated to AI chip orders secured by Intel but to “any customer that we could potentially see some business from who’s expressed some interest — we’re counting that then has to be converted to real dollars.”

“There’s a need for GPUs to do that [AI] work. I think we are a beneficiary of that because of the CPU that we have,” Zinsner said.

Many Nvidia H100 GPU installations are alongside Intel’s Sapphire Rapids chip, which supports DDR5 memory. However, a larger amount of dollars are going to GPUs for large language models, which will continue to hurt Intel’s revenue in the coming quarters.

“That will really be our story in 2024 — driving Gaudi together with CPUs. Ultimately, we’ll have our Falcon Shore GPU product out in 2025. A lot of that is building the software ecosystem,” Zinsner said. Intel is trying to reduce the confusion around its fastest AI chips by merging the discrete Gaudi AI chips into the Falcon Shores GPU.

“What will happen is Gaudi will converge with Falcon Shores,” Zinsner said, adding, “There will be one product offering.” That will pack Intel’s fastest AI chips into a single product offering, reducing the confusion on whether customers should build their AI computing around Gaudi or Falcon Shores.

The general-purpose Falcon Shores with built-in Gaudi features are being positioned as a competition to Nvidia’s GPUs. Like the H100, Falcon Shores supports general computing but also has low-precision compute functions that specialize in drawing conclusions from patterns and trends in data.

The Gaudi lineup is also targeted at customers looking to build AI rigs in-house or in controlled environments and not in the cloud, Zinsner said.

Intel Gaudi 2 AI Processor (Source Intel)

“As you look beyond the really big parameter training that requires scrubbing the whole Internet, and you start to look at it in a more contained environment, particularly in on-prem searches, it has definitely got a real performance level that is equivalent to what you see from competitors,” Zinsner said.


Part of what is not said here is interesting to me. Intel had two product lines competing for AI sockets. They had the Ponte Vecchio done by R. Koduri and team, which is a product similar to the AMD MI250 or MI300x. They also had the Gaudi product line done by the acquired Habana corp. R. Koduri left Intel and the successor to the Ponte Vecchio was canceled. The next follow on known as Falcon shores has been delayed until 2025, and has been redefined as having no CPU tiles, which is similar to the change AMD made in going from the MI300A to the now AI centric MI300x.

Intel has stated Gaudi is selling much better into AI sockets than PVC, but PVC was really designed as a supercomputer accelerator like the MI250 and MI300a. The currently shipping Gaudi 2 which the article mentions is on TSMC N7 still, while the Gaudi 3 due out next year is on TSMC N5 like the PVC, or AMD MI300. I suspect Falcon Shores is really just Gaudi 4, and will likely be on Intel_18A.

Intel is somewhat famous for setting up internal competitions and running with the winner. It appears to me Koduri lost and Habana won.

P.S. Interesting aside that TMF does now allow you to post Koduri’s first name.


Any thoughts on how the shift to a separate business unit for FPGA/Altera dovetails with this?

This is more similar to what they did with Mobileye. They continue to pare back so they focus on their main thrust of X86 and foundry.