Intel has officially unveiled Crescent Island, its next-generation AI inference accelerator built on the Xe3P architecture. This specialized GPGPU lacks traditional 3D graphics blocks, focusing instead on air-cooled data center and workstation solutions.
The hardware features 32 Xe3P cores across four compute slices, totaling 256 Vector and XMX engines. It supports data formats from FP4 to FP64 and includes 32 MB of L2 cache. Reference PCIe cards will offer 160 GB of LPDDR5X memory with a 350W TDP rating.
By utilizing the same architecture teased for Panther Lake and future Arc C-Series products, Intel is streamlining its AI stack. The company expects to begin customer sampling in the second half of 2026, targeting high-density inference and GPGPU workloads.