Intel Details Crescent Island GPU With 480GB Memory For AI Data Centers
Intel is positioning Crescent Island as an AI-first server GPU built around power efficiency and KV cache capacity rather than peak throughput, a bet that its Arc Pro B-series lineage can be redirected at agent workloads while SambaNova's SN50 fills the Xeon pairing role.
Reporting from 1 source: ASCII.jp.
Intel's Crescent Island is an AI data center GPU built on the Xe3p architecture, described at Hot Chips as targeting agent-style AI workloads and the KV cache problem. It is a 32 Xe core part with five Xe3p cores per Compute Slice across four slices. Intel gives a 350W TDP, keeps air cooling, and adds native MxFP4 support using the OCP block size 32 definition without custom extensions. Shipments were discussed for the second half of 2026.
Crescent Island keeps the consumer Xe lineage rather than the Xe-HPC line: eight Vector and eight XMX units make up one Xe core, grouped five cores to a Compute Slice, four slices to the part. The Xe3p additions are FP4 support and a 16x16 systolic array inside XMX, which Intel did not extend beyond OCP's block size 32 MxFP4 format. Intel set a 350W TDP and stressed that air cooling remains the target, framing the design around performance per watt instead of absolute throughput.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.