Intel on Thursday unveiled three additions to its Intel Xeon 6 series of central processing units (CPUs) that are designed to manage graphics processing unit (GPU) powered systems. One, the Xeon 6776P, is currently serving as the host CPU for Nvidia’s DGX B300, its latest generation of AI-accelerated systems.
According to a release from Intel, the Xeon 6776P processor “plays a vital role in managing, orchestrating and supporting the AI-accelerated system. With robust memory capacity and bandwidth, the Xeon 6776P supports the growing needs of AI models and datasets.”
It added that the new processors offer Performance-cores (P-cores) that include Intel’s Priority Core Turbo (PCT) technology and Intel Speed Select Technology – Turbo Frequency (Intel SST-TF), which delivers customizable CPU core frequencies to boost GPU performance across demanding AI workloads. They also boast up to 128 P-cores per CPU, up to 30% faster memory speeds compared to the competition, up to 20% more PCIe lanes than previous models, and support for FP16 precision arithmetic.
Intel said that the introduction of PCT, paired with Intel SST-TF, marks “a significant leap forward in AI system performance. PCT allows for dynamic prioritization of high-priority cores (HP cores), enabling them to run at higher turbo frequencies.”
It added, “in parallel, lower-priority cores (LP cores) operate at base frequency, ensuring optimal distribution of CPU resources. This capability is critical for AI workloads that demand sequential or serial processing, feeding GPUs faster and improving overall system efficiency.”
Jeremy Roberts, senior director of research and content at Info-Tech Research Group, described the announcement from Intel about its inclusion in the Nvidia DGX B300 as “really a tale of two stock charts. Intel (massive declines) and Nvidia (massive gains) have competed in various forms for decades, but with AI and LLMs and the demand for data center chips that can meet AI’s performance requirements, it represents a moment for cooperation.”
He added that his read is that “Intel recognizes that Nvidia is far and away the leader in the market for AI GPUs and is seeking to hitch itself to that wagon.”
Roberts said, “basically, Intel, which has struggled tremendously and has turned over its CEO amidst a stock slide, needs to refocus to where it thinks it can win. That’s not competing directly with Nvidia but trying to use this partnership to re-secure its foothold in the data center and squeeze out rivals like AMD for the data center x86 market. In other words, I see this announcement as confirmation that Intel is looking to regroup, and pick fights it thinks it can win. “
He also predicted, “we can expect competition to heat up in this space as Intel takes on AMD’s Epyc lineup in a push to simplify and get back to basics.”
Matt Kimball, vice president and principal analyst, who focuses on datacenter compute and storage at Moor Insights & Strategy, had a much different view about the announcement.
The selection of the Intel sixth generation Xeon CPU, the 6776P, to support Nvidia’s DGX B300 is, he said, “important, as it validates Intel as a strong choice for the AI market. In the big picture, this isn’t about volumes or revenue, rather it’s about validating a strategy Intel has had for the last couple of generations — delivering accelerated performance across critical workloads.”
Kimball said that, In particular, there are a “couple things that I would think helped make Xeon the chosen CPU.”
These include the ability to selectively deliver performance boosts through its PCT and SST-TF technologies “to enable a faster feeding of GPUs,” and support for MRDIMM (multiplexed rank DIMM) memory that can access two memory DIMMs simultaneously.
This parallelism, he said, “effectively allows twice the data throughput (128 bytes versus 64 bytes). MRDIMM also allow faster transfers (MT/s) which further removes the bottleneck.”
And, he added, “while I think the AMX (Advanced Matrix Extensions) AI accelerator built into Xeon is a great differentiator for the company, I don’t believe this had a play in Nvidia’s selection simply because it helps Xeon execute matrix operations in support of AI for training and inference on Xeon. While I’m certain that it could provide some relief in terms of pre-processing, there is not a clear statement in how Xeon is being utilized. So, perhaps complementary to some degree. “
Kimball said, “overall, I think it’s the ability to squeeze better performance out of cores, and faster memory access, and bandwidth are the reasons for the selection. And I believe this selection is important for Intel, as Nvidia has effectively signaled to the enterprise that it sees Intel as the optimal x86 controller for its AI clusters.”
Source:: Network World