Intel Crescent Island Packs 480 GB LPDDR5X for AI Inference Workloads

A GPU That Cannot Draw a Single Pixel

Intel took the wraps off Crescent Island at Hot Chips 2026, and the spec sheet reads like something built for a very specific, very expensive job. According to TechPowerUp, the card is based on the Xe3P GPU IP and runs at a 350 W thermal envelope in a standard air-cooled, PCIe form factor that fits any compatible configuration. There are 32 Xe3P cores onboard, 256 XMX engines, 32 MB of unified L2 cache, and a baseline memory load of 160 GB of LPDDR5X on Intel’s own branded cards.

That last number is already enormous by any consumer standard, but Intel is letting ODM partners push the configuration further. Those partners can equip Crescent Island cards with up to 480 GB of LPDDR5X memory, which puts the potential aggregate capacity of a four-card workstation somewhere in the neighborhood of 1.92 TB. As TechPowerUp notes, running a trillion-parameter model like Kimi K3 locally requires nearly 2 TB of system memory. Intel’s math checks out, if not exactly your budget.

Oh, and here is the thing: Crescent Island cannot display anything. The Xe3P silicon includes no graphics processing elements for image output. It is, as TechPowerUp reports, designed purely for AI inference workloads including long-context and agentic tasks. The “graphics card” label is largely a form-factor courtesy at this point.

The Architecture Behind the Numbers

What makes Xe3P distinct from the standard Xe3 used in consumer Arc graphics is the depth of its XMX engine redesign. Where regular Xe3 uses a four-deep systolic design, Xe3P steps that up to 16-deep processing, according to TechPowerUp. That shift has meaningful implications for how efficiently the hardware handles matrix math at scale, which is the core workload Intel is targeting.

Each of the 32 Xe3P cores carries a 1 MB register file and 512 KB of L1 cache. The full card totals 256 Vector engines and 256 XMX engines, organized in groups of eight per core. Intel also added FP4 data format support alongside FP8, giving the card more flexibility in how it handles lower-precision inference tasks. That kind of format breadth matters a lot when you are trying to squeeze trillion-parameter models through hardware as efficiently as possible.

Intel is also promising day-zero support for all major AI frameworks, with software integration already in place at launch. That is a practical detail worth noting for workstation buyers who have been burned before by driver and framework compatibility headaches.

Who Actually Buys This

Intel is targeting three deployment categories with Crescent Island: PC, workstation, and data center edge. The workstation pitch is perhaps the most compelling to think about concretely. Four cards, each carrying the maximum 480 GB configuration, on a single machine capable of running frontier open-source models locally. The caveat TechPowerUp flags is real and hard to ignore: DRAM is expensive, and 480 GB of LPDDR5X per card is not going to come with a budget-friendly price tag.

Still, for AI researchers, enterprise edge deployments, or labs that want to keep model inference off cloud infrastructure entirely, the option exists. The combination of massive memory capacity, a standardized PCIe form factor, and ready framework support gives Crescent Island a practical argument that is harder to dismiss than the raw spec numbers might suggest on their own.

Whether the price of those four-card configurations ends up making local trillion-parameter inference genuinely accessible, or remains the kind of thing only well-funded teams can justify, will be answered when actual pricing arrives.