Intel Crescent Island Specs: 480GB Of LPDDR5X On One Air Cooled Card

480GB of memory on a single accelerator. That is what Intel detailed at Hot Chips 2026, and it got there by ignoring the component everyone else is fighting over.

Crescent Island ships with 160GB of LPDDR5X as standard, with ODM designs supported up to 480GB. It uses no HBM4 and no liquid cooling, and runs at a 350W TDP on an air cooled PCIe Gen5 x16 card.

The Crescent Island specs that matter

The part carries 32 Xe3P cores, 256 XMX engines and 32MB of unified L2. The architecture work underneath is not cosmetic. The XMX systolic array went from four deep to sixteen deep, and the Xe Vector Engine picked up FP8, FP4 and microscaling formats. The general purpose register file doubled to 1MB per Xe core, up from 512KB on Battlemage, with a further 512KB of L1 or shared local memory per core.

Those are Intel's published figures. Nobody has run independent inference benchmarks against them, so treat every number as vendor stated until someone does.

Why the LPDDR5X choice is the whole story

Read it next to the two racks that launched around it and 2026 gets easier to parse. AMD Helios and Nvidia Vera Rubin are both HBM4, both liquid cooled, both built for bandwidth per watt, and we put those two side by side in our Helios versus Vera Rubin breakdown. Intel went the other way, with cheaper memory in much larger quantities inside a power envelope that fits a rack that already exists.

That is a bet about where the bottleneck sits. If serving a model is memory bound rather than compute bound, how much of it you can hold matters more than how fast you can multiply. HBM4 exists because bandwidth is expensive, and we covered what HBM4 actually buys you when the AMD rack landed. Intel is not trying to win that comparison. It wants to be the card you drop into an air cooled datacentre without rebuilding the plumbing.

The trade is real. LPDDR5X bandwidth does not come close to HBM4, so Crescent Island will lose badly on anything bandwidth limited. Intel appears to have accepted that in exchange for capacity and cost.

What to watch

The number that decides this is cost per gigabyte at the rack level, not FLOPS, and Intel has not published pricing. That matters more than usual right now, because DRAM contract prices have roughly doubled twice this year, the same squeeze behind the consumer memory crisis. Commodity LPDDR5X is not automatically the cheap option any more.

Until Intel prices it, and until someone benchmarks all three parts on the same inference workload, 480GB is a specification rather than an advantage.