# Intel Crescent Island Skips HBM: 480 GB of LPDDR5X on a 350 W Air-Cooled Card > Source: > Published: 2026-08-25 12:12:08+00:00 *Up to 480 GB of laptop-class memory on a datacenter accelerator, and Intel is treating the absence of HBM as the selling point.* Intel used its [Hot Chips 2026 session](https://newsroom.intel.com/client-computing/intel-outlines-architectures-for-agentic-ai-at-hot-chips-2026) to fill in the blanks on **Intel Crescent Island**, the Xe3P inference accelerator it first teased last autumn. The headline number is one nobody else in the datacenter GPU business can currently print: up to 480 GB of memory on a single PCIe add-in card. The catch, or the whole point depending on where you sit, is that none of it is HBM. Crescent Island runs LPDDR5X, the same class of memory soldered next to the CPU in a thin-and-light laptop, just an absurd amount of it. Intel’s own reference card carries 160 GB; board partners can populate the design out to 480 GB. Behind the memory sit [32 Xe3P cores, 256 XMX matrix engines](https://www.techpowerup.com/351901/intel-details-crescent-island-graphics-32-xe3p-cores-up-to-480-gb-lpddr5x-memory), 32 MB of unified L2 and a PCIe Gen5 x16 link, all inside a 350 W air-cooled card. Read that last part twice. Three hundred and fifty watts, air-cooled, standard slot. The accelerators everyone is currently fighting over need liquid loops, custom trays and a rack refit before they will even power on. Crescent Island is built to drop into infrastructure that already exists. ## Capacity is the constraint now, not FLOPS The reasoning is easy to follow if you have tried to serve a large model recently. Weights, KV cache and long contexts all eat capacity, and the moment a model stops fitting in one device you start sharding it across several and paying an interconnect tax on every token. A single card holding 480 GB removes that problem for a whole class of models, with no fabric, no NVLink-style plumbing and no second server. Bandwidth is what pays for it. [VideoCardz reckons](https://videocardz.com/newz/intel-crescent-island-gpu-to-support-lpddr5x-9600-memory-and-1-5-tb-s-bandwidth) an LPDDR5X-9600 configuration on a very wide bus lands somewhere around 1.5 TB/s, which is respectable for LPDDR and nowhere near what a modern HBM stack delivers. Intel is not pretending otherwise. Crescent Island is pitched at inference and agentic workloads, not training, and the company has been careful not to line it up against the top of Nvidia’s stack. ## Opting out of the HBM queue There is a supply-chain argument underneath the technical one, and it may end up mattering more. HBM is the industry’s chokepoint: allocation is spoken for years ahead, and memory has swollen into a serious fraction of what an AI rack costs. LPDDR5X is manufactured in phone and laptop volumes and is cheap per gigabyte. By building around it, Intel sidesteps the auction entirely, which is an elegant way of turning a weak negotiating position into a product. Whether that converts into sales is another question, and Intel’s record here invites caution. Ponte Vecchio arrived late and left quietly, Falcon Shores was demoted to an internal test vehicle before it ever shipped, and Gaudi never won the software traction it needed. Silicon has rarely been the problem in this segment; the stack around it has. Customer sampling is slated for the second half of 2026, so there are no independent performance numbers yet, and until somebody runs real models on real cards, 480 GB is a spec sheet rather than a result. It is still the first genuinely differentiated pitch Intel has made in accelerators in years. Everyone else is selling bandwidth. Intel is selling room.