Intel details Crescent Island data center GPU with 480 GB of LPDDR5X

What's new

At Computex in Taipei on June 1, 2026, Intel disclosed additional technical detail on Crescent Island, its next-generation data center GPU aimed at inference. The part is built on the Xe3P architecture, an extension of the Xe line that Intel positions for higher performance-per-watt while keeping software compatibility with its existing stack. The headline figure is memory: Crescent Island uses LPDDR5X rather than HBM or GDDR, and scales to as much as 480 GB of capacity per card. Intel frames that choice as a way to hold large, token-intensive context windows close to the chip and lower total cost of ownership, trading the higher bandwidth of HBM for far greater capacity and lower power. The card is a 350-watt air-cooled PCIe add-in design rather than an OAM module. On datatypes, Intel says Crescent Island adds native support for FP4 and MXFP4 microscaling formats and spans up to FP64, broadening the range of operations the silicon handles in hardware. Intel says developers can build and validate on its existing Arc Pro Series GPUs, which share the Xe foundation, and deploy to Crescent Island with forward and backward compatibility. Customer sampling is planned for the second half of 2026. Source: Intel Newsroom, "Intel Puts Agentic AI to Work with Xeon 6+, Networking, and AI Systems," June 1, 2026.

Why it matters

Crescent Island is Intel's clearest data center accelerator pitch since the Gaudi line, and the memory strategy is the differentiating bet. With HBM supply tight and inference workloads increasingly bound by how much model state and context a single card can hold, 480 GB of LPDDR5X reframes the question from peak bandwidth to capacity per watt. A 350-watt air-cooled PCIe envelope also lets the card drop into standard servers without liquid cooling or OAM baseboards, which lowers the barrier for enterprises that want local inference without rebuilding racks. Native FP4 and MXFP4 support aligns Crescent Island with the low-precision formats that NVIDIA and AMD have moved to for inference throughput. The same Computex briefing paired the GPU with Xeon 6+ on Intel 18A, signaling that Intel intends to sell the host CPU and the accelerator as a coupled system rather than competing on the GPU alone.

Caveats

Intel disclosed architecture and capacity but not the figures that decide competitiveness. There is no published memory bandwidth, which is the central trade against HBM parts, and no peak or sustained throughput numbers at any precision. No MLPerf or third-party benchmark accompanies the disclosure, so performance against NVIDIA and AMD inference cards is unverified. The 480 GB figure is the top configuration; a standard reference card is reported at 160 GB, and Intel has not detailed the SKU range or pricing. Sampling is set for the second half of 2026 with general availability later, so timelines could move. As with any vendor disclosure, the framing is Intel's, and the capacity-over-bandwidth argument has not been tested against real agentic-inference traces.