Intel Preps Diamond Rapids and Crescent Island for the AI Agent Era

Intel used Hot Chips 2026 to show how it wants to split AI workloads across processors, accelerators, and end-user devices. The company detailed Diamond Rapids, its next-generation Xeon with up to 256 cores; Crescent Island, an inference GPU with up to 480 GB of memory; and Wildcat Lake, already shipping as the Intel Core Series 3 — three different architectures meant to cover everything from large data centers to laptops and edge systems.

Intel’s new AI architecture: the key points in 30 seconds

• Diamond Rapids will scale up to 256 cores, 16 memory channels, and 128 PCIe 6.0 lanes with CXL 3.0.
• Crescent Island packs 480 GB of LPDDR5X onto a 350 W, air-cooled PCIe card.
• The GPU is designed specifically for inference and for running more agents at once.
• Wildcat Lake, the Intel Core Series 3, brings this strategy to laptops and the edge with an NPU rated up to 17 TOPS.
• Intel is using 18A and 18A-P, Foveros Direct 3D, and UCIe as part of this heterogeneous architecture.

The approach is interesting because Intel isn’t trying to present a single chip capable of covering every AI need. Its proposal starts from the opposite premise: agents require different types of compute depending on where they run and what they’re doing.

A server may need lots of cores to coordinate processes, memory, and accelerators. Token generation needs a GPU with enough memory capacity and bandwidth. And a laptop needs to run certain functions locally without spiking power consumption.

Intel wants a presence at all three layers.

Diamond Rapids Pushes Xeon to 256 Cores

Diamond Rapids will be one of the major changes in the next generation of Intel Xeon processors.

The company has confirmed configurations with up to 256 cores, paired with a massive last-level cache (LLC) of up to 1.28 GB. On top of that, there are 16 memory channels capable of running at 12,800 MT/s and 128 PCI Express 6.0 lanes compatible with CXL 3.0.

The processor will be built on Intel 18A-P, an evolution of Intel 18A aimed at improving performance and thermal characteristics. Intel Foundry confirmed in June that 18A-P had entered risk production.

Diamond Rapids will also use Foveros Direct 3D and UCIe-S to interconnect the processor’s different pieces.

Here’s another one of Intel’s important bets.

Large processors are increasingly moving away from behaving like monolithic pieces of silicon. Using chiplets allows specialized blocks to be combined within the same package, and UCIe (Universal Chiplet Interconnect Express) aims to establish a standard for connecting them.

Intel will also add new Advanced Performance Extensions (APX) and improved versions of Advanced Matrix Extensions (AMX), instructions aimed, among other uses, at accelerating the matrix operations common in AI.

But Diamond Rapids isn’t meant to replace a GPU for running large models.

Its role within Intel’s proposed architecture is different: providing general-purpose capacity to feed, coordinate, and manage platforms where CPU, memory, storage, network, and accelerators all have to work together.

This matters especially in agentic systems.

An AI agent isn’t necessarily limited to generating text. It can query databases, run code, call APIs, access storage, perform searches, and coordinate other agents. Part of that work depends more on CPU, memory, and I/O than on heavy matrix compute capacity.

Crescent Island Bets on 480 GB and Just 350 W

Of the three, Crescent Island is probably the most eye-catching proposal.

Intel is building this GPU specifically for inference and has opted for an unusual combination: lots of memory and relatively contained power consumption.

Crescent Island will pack 32 Xe cores and 256 XMX engines based on Xe3P, paired with up to 480 GB of LPDDR5X memory.

All of that will be packed into a PCIe card with 350 W power consumption that can be air-cooled.

The memory figure matters a lot.

A considerable part of the cost of running AI models doesn’t come solely from performing math operations. Model weights have to be stored in memory, and as context grows or the number of simultaneous users and agents increases, pressure on that memory grows too.

Intel is trying to tackle that problem through capacity.

The 480 GB would allow large models to be hosted, or a larger number of concurrent processes to be served, without necessarily resorting to multi-card configurations purely to get enough memory.

The company is using LPDDR5X instead of high-bandwidth HBM memory. It’s a decision that points directly at the balance between capacity, power consumption, cost, and inference performance.

Crescent Island shouldn’t be compared to the market’s largest accelerators on peak power alone. Intel presents it as a GPU designed to improve the economics of token generation inside data centers that already use conventional air cooling.

The 350 W figure matters precisely for that reason.

The highest-performing AI GPUs are rapidly increasing their power and thermal demands. Installing very high-power accelerators can force changes to racks, power delivery, and even cooling systems.

Crescent Island tries to fit into more conventional infrastructure.

Intel maintains that this combination will allow larger models to run, wider context windows to be used, and more simultaneous agents to be supported. These are design goals stated by the company, and independent testing will be needed to find out its real performance per token and compare it against alternatives from NVIDIA, AMD, and other vendors.

That figure will matter quite a bit more than the 480 GB on its own.

Large memory capacity allows bigger workloads to be hosted, but final performance also depends on bandwidth, compute capacity, the precision used, software, the model, and the level of concurrency.

Wildcat Lake Brings the Same Idea to the Laptop

The third piece is already much closer to the end user.

Wildcat Lake is the codename for the Intel Core Series 3, a generation aimed at laptops and edge platforms where price and power consumption carry considerably more weight than in a data center.

Intel has confirmed configurations with two performance cores and four efficiency cores, integrated Xe3 graphics with XMX acceleration, and a neural processing unit (NPU) capable of reaching 17 TOPS for hybrid AI workloads.

It also supports LPDDR5X-7467 memory, Wi-Fi 7, and Bluetooth 6.0.

Wildcat Lake is also interesting for another reason: it’s the first Intel processor to use UCIe.

Intel’s technical documentation shows that the Core Series 3 uses a multichip architecture with different manufacturing technologies. The processor combines Intel 18A with N6 and N3E processes within the same product.

It’s a good demonstration of where the company wants to head.

Intel 18A doesn’t necessarily mean absolutely every piece of a processor has to be built on 18A. Chiplets allow different processes to be chosen based on each block’s needs, and then joined together within the package.

The strategy also fits with hybrid AI.

It doesn’t make much sense to send every operation to a data center if certain tasks can be handled on the device itself.

Speech recognition, image processing, video-call effects, classification, small models, and certain assistant functions can run locally. More demanding operations can continue to be offloaded to servers.

Intel calls this approach Hybrid AI.

The challenge will be getting software and applications to transparently make use of the CPU, GPU, NPU, and remote accelerators depending on each task.

The company is thus sketching out an infrastructure where Diamond Rapids provides general-purpose compute on the server, Crescent Island primarily handles inference, and Wildcat Lake runs part of the AI close to the user.

That doesn’t mean all three products have to be used together. They’re architectures aimed at different markets.

But together, they show a fairly clear strategy.

Intel believes the growth of AI agents won’t depend solely on building ever-bigger accelerators. It will also require CPUs capable of coordinating huge platforms, inference cards that can be installed without completely rebuilding the data center, and processors capable of handling part of the workload locally.

That vision will still have to face the market, though.

NVIDIA currently dominates much of the infrastructure used to train and run models, AMD is growing its presence in accelerators, and the big cloud providers are developing their own specialized chips.

Intel needs to prove performance, availability, and, especially, software capable of making its alternative attractive.

Crescent Island will probably be one of the most interesting tests.

480 GB of memory on a 350 W, air-cooled PCIe card is a very different proposition from the race to build the accelerator with the highest power draw and peak performance.

If it manages to translate that architecture into a good cost per token, Intel could find an interesting niche in a market that’s starting to care as much about how much it costs to run AI as about how much peak performance it can reach.

Frequently Asked Questions

How many cores will Intel Diamond Rapids have?

Intel has confirmed Diamond Rapids configurations with up to 256 cores, 1.28 GB of LLC cache, 16 memory channels at 12,800 MT/s, and 128 PCIe 6.0 lanes compatible with CXL 3.0.

How much memory will the Intel Crescent Island GPU have?

Crescent Island will be able to pack up to 480 GB of LPDDR5X memory. Intel designed it as a 350 W, air-cooled PCIe card aimed primarily at AI inference.

Is Crescent Island designed for training or inference?

Intel presents Crescent Island specifically as a data center GPU optimized for inference, with particular attention to the number of tokens processed, memory capacity, and running multiple agents simultaneously.

What is Intel Wildcat Lake?

Wildcat Lake is the codename for the Intel Core Series 3 processors. They use Intel 18A alongside other processes in a multichip architecture and include Xe3 graphics and an NPU rated up to 17 TOPS for running certain AI workloads locally.