Intel on Tuesday launched Xeon 6 P-core chips for data center CPU needs alongside Gaudi 3 accelerators for AI, with both designed to reduce power consumption and costs to enterprise users while bolstering security and offering greater performance over previous generations.
The updates come after a series of strategic corporate maneuvers designed to improve Intel’s lagging fortunes, but analysts reacted skeptically, questioning if either chip can do much to improve the longtime chipmaker’s downward trajectory any time soon.
“Intel won’t be able to reverse its challenges in the short term,” even with Xeon 6 and Gaudi 3, said analyst Jack Gold of J. Gold Associates. “It will likely take two to three years for them to fully recover.”
A broad industry transition to inference AI as opposed to high end model training will be needed to boost Intel in the data center market, Gold said.
“Once the market shifts to 75%-plus of workloads in AI becoming production inference solutions, the enhanced Xeon and Gaudi systems become much more relevant,” Gold said.
IDC analyst Mario Morales said the new Xeon and Gaudi chip launches do not make him feel better about Intel’s financial and technology outlook, which took a nosedive after the last earnings call Aug. 1 when 15,000 layoffs were announced. “I see a lot more pain the company has to go through to overcome the challenges they’ve self-inflicted,” Morales said.
One key negative with the new Gaudi 3 silicon is that competitor TSMC will manufacture the complex Intel design, even though the Intel Foundry is where it should have ideally been produced, Morales said. (Intel CEO Pat Gelsinger said a week ago that Intel plans to establish Intel Foundry as an independent subsidiary inside Intel.)
The Xeon 6 announcement allows Intel to compete, Morales said, “but it doesn’t put Intel over the top; it’s still playing catch up to AMD on the CPU side.”
Such criticisms aside, Intel executives spent a day with reporters and analysts at an embargoed presentation in Oregon describing the power and cost efficiency and performance virtues of the new chips.
A raft of customers were also announced, and executives from IBM and Dell were on hand to say how they will deploy Gaudi 3 in products for customers, even as both companies will still rely on AI chips from Nvidia and AMD.
“With Gaudi 3, what caught our eye was significant cost-performance benefit. We’ve been co-engineering in the early phase and that led to the announcement we’ll be rolling out some of this in our MZR fleets and going after joint enterprise clients,” said Rohit Badlaney, general manager of IBM cloud product and industry platforms.
Vivek Mohindra, senior vice president of corporate strategy at Dell, said the company’s flagship 9680 server will be launched with Xeon and Gaudi 3. “When you think of enterprises, they really do need that easy button to deploy it, but have confidence it’s going to result in an ROI for them,” Mohindra said.
In a video endorsement, Google Cloud general manager Mark Lohmeyer said the company will use Xeon 6 processors for its cloud customers, having seen the value of previous Xeon generations.
In a slide, Intel named a number of companies that are Gaudi and Xeon customers, including Meta, Ubuntu and Seekr.For Gaudi 3, Intel announced the air-cooled Dell PowerEdge XE9680 will ship in October.
Tech specs for Intel Gaudi 3, Xeon 6
Gaudi 3 will come in variants, including an accelerator card, universal baseboard and a PCIe HL-338 Add-in card. The latter will feature 1835 TFLOPS and eight matrix multiplications engines. With four cards per system, Intel described it as “ideal for inferencing, fine tuning and small model training.”
Compared to H100 GPUs from Nvidia, Gaudi 3 has 1.09 times greater inference throughput running LLaMA 3 and 1.8 times the performance per dollars of inference throughput, Intel said.
Morales said he regards Gaudi 3 as a “stopgap” for Intel’s roadmap and competitive value. “Intel seems to be building momentum but it’s no where near where incumbents are today,” he said. In terms of long term strategy, he is wondering what Intel plans after Gaudi 3 with the code-named Falcon Shores, a combination of Hibana IP and a GPU. “What happens after Gaudi 3, but they’ve not talked about that at all and because they didn’t emphasize it tells me it won’t help Intel as much as initially though about what puts them on the GPU map,” Morales said.
“By not talking about Falcon Shores, it tells me they will put more focus on inferencing jobs and Intel will miss out on high end frontier models like Open AI,” Morales said. ”They are saying that high end modeling will stay with Nvidia and AMD, with AWS and Google implementing it themselves, while Intel focuses on the endpoints at the edge and inferencing.”
Intel recently announced a big deal with AWS to produce custom AI and Xeon 6 chips as well as a $3 billion government-focused Secure Enclave project , but Morales said both will have impact long term—in five to six years—and “won’t help in the next year.”
RELATED: Intel lifted by AWS commitment to 18A and deal for custom chips
Xeon 6 P-series chips focus on performance needs of data centers, while earlier announced Xeon 6 E-series refers to efficient performance for when top performance is not needed in a compute task.
Intel executives said the Xeon 6 P-cores are still designed to handle compute-intensive workloads with “exceptional efficiency” with double the performance of the prior generation. There are more cores and double the memory bandwidth and AI acceleration in every core. Intel recorded up to double the GPT-J-6B performance with the Xeon 6980P versus a prior generation. Also, Intel said the Xeon 6900 series (part of 6th generation Xeon) has up to 3.7 AI inference performance compared to AMD EPYC. And, integer throughput with 6980 is nearly 6 times higher than and a prior generation.