Intel hopes its latest chips will resolve AI-related power efficiency and cost concerns even as rivals AMD and Nvidia turn up their power consumption to deliver higher performance.
The company introduced a flagship Xeon 6+ CPU, a new GPU called Crescent Island, and a new Ethernet controller on Monday.
These are the first major data-center releases since Lip-Bu Tan became CEO in March last year.
Intel, once a byword for chip innovation, is shifting its philosophy to help customers balance systems for cost and performance, as opposed to selling the most powerful chips, said Kira Boyko, product line director for Xeon products.
“Our customers constantly tell us this is becoming too expensive and taking up too much energy. So, we look at that full view from the very beginning of product design,” Boyko said.
Intel’s new offerings should enable significant server consolidation and lower floor space requirements while increasing throughput, said Stephen Sopko, analyst-in-residence at Hyperframe Research.
“Explosive AI growth has created major constraints around power, cooling, and operational costs,” Sopko said.
A GPU for Agentic AI
Intel announced a new GPU called Xe3P, code-named Crescent Island, which “is purpose-built for this upcoming AI generation of agents,” said Anil Nanduri, vice president of AI products at Intel’s Data Center Group.
“We are focusing on cost efficient inferencing for the data center,” he said.
Agentic AI workloads involve humans spawning hundreds of thousands of agents, and the 350-watt Crescent Island GPU can run multiple expert agents, coordinating actions across CPUs or other accelerators, he said.
“There is no traditional graphics or 3D support. In fact, removing some of these features enabled us to give you more area and silicon to increase the AI performance,” Nanduri said.
Intel opted for 480GB of LPDDR5X memory — typically used in PCs and smartphones — for its low cost and power efficiency.
“We made a very deliberate design choice of using LPDDR memory instead of GDDR: LPDDR5X memory with densely packed channels delivers significant bandwidth and most importantly, memory capacity,” Nanduri said.
Memory supply constraints have driven up memory and GPU prices. A Trendforce report last month said LPDDR5X demand is tightening capacity. The memory type has access to “a different set of memory partners” to keep the supply rolling, Nanduri said.
Nvidia and AMD pack another graphics memory type, HBM (high-bandwidth memory), in their fastest GPUs. HBM is supplied predominantly by three vendors, SK Hynix, Samsung and Micron, which hold 95% of market share, according to a Counterpoint Research study. Trendforce’s report also noted HBM facing a 100% price rise in 2027 due to shortages.
A shipment date for Crescent Island will be announced later, Nanduri said.
Intel also announced the Xeon 6+, code-named Clearwater Forest, which was delayed from last year. It is the first chip on Intel’s 18A manufacturing process, with new transistor features making it faster and more power efficient.
As a chip, the CPU has reemerged from under the shadow of the GPU as it handles AI agents and distributes workloads among accelerators, Boyko said.
The missing link
“We’re actually seeing a large amount of GPUs that are really underutilized because they don’t have the CPUs to support them and help with… the orchestration of workloads,” Boyko said.
Xeon 6+ offers 288 low-power E-cores, up to 8,000 megatransfers per second of DDR5 memory and 576 megabytes of last level cache.
There were many design changes in Clearwater Forest through the recent leadership changes and AI’s impact on computing requirements, Boyko said.
For example, the Xeon 6+ was designed to initially support 7200 megatransfers per second, but customers wanted faster bandwidth to handle AI orchestration and agents.
Xeon 6+ includes AET (Application Energy Telemetry), a new feature that gives operators real-time visibility into energy use at the application level. That allows for better orchestration and tuning of workloads, Boyko said.
It also lets data-center providers bill customers accurately based on actual energy use rather than estimates, Boyko said.
This allows for accurate charge backs, or for operators to offer rebates, “giving an incentive to an end customer to orchestrate differently or use energy differently so that they’re reducing their overall costs as well,” Boyko said.
Major OEMs will announce Xeon 6+ servers in June, Intel said. Server makers will also plug the new chip into systems running Xeon 6 CPUs because of socket compatibility, Boyko said.
Security over speed
Intel also introduced the E835 Ethernet card. At 200Gbps it isn’t the fastest silicon as network cards are already achieving 400Gbps, but Intel highlighted its security features and power efficiency.
E835 consumes 47% less power than Nvidia’s ConnectX6 and 28% less than Broadcom BCM957508-P2100G at full 200Gbps bidirectional load, said Brian Neipoky, who is the Ethernet product line senior director at Intel, during the media briefing.
Jack Gold, principal analyst at J. Gold Associates, said Intel has a good strategy in mind as billing for AI tokens skyrockets and power use with token generation is high.
Monitoring power and cost can reduce data center costs but also predict failures and extend chip lifecycles. “Along with high temperature, high power usage is not often kind to chip longevity,” Gold said.
Challenges remain. Intel doesn’t have a coherent AI strategy to connect its PC, data center and edge markets, Gold said. Over the last few years, Intel has conceded the GPU market to Nvidia after cancelling multiple GPU products under former CEO Pat Gelsinger.
But it also needs to regain lost ground in CPUs, he said. Hyperscalers Google, Amazon, Meta and Microsoft have developed their own ARM-based CPUs, and AMD is gaining share in the x86 market.
“The question is, will x86 remain the dominant player in that space as others, mostly ARM-based, try to capture market share?” Gold said.