Intel Draws First Blood at Computex: Crescent Island GPU Packs 480GB Without HBM

📡 AI News 💬

🩺 Summary

Intel unveils Crescent Island GPU at Computex: up to 480GB LPDDR5X memory — no HBM needed. 350W air-cooled, targeting agentic AI inference. Sampling H2 2026.

📝 Details

Everyone in the AI industry knows HBM is in critically short supply, but Intel chose a different path entirely. At Computex 2026, CEO Lip-Bu Tan personally unveiled Crescent Island — Intel's next-generation data center GPU. Its biggest selling point isn't raw compute or architecture, but 480GB of LPDDR5X memory. You read that right: LPDDR5X — the same memory chips found in laptops. While everyone else is fighting tooth and nail over HBM allocation, Intel is saying: skip HBM, use LPDDR5X, and pack in more capacity. It's a fascinating bet. Built on Intel's Xe3P GPU architecture, Crescent Island is officially positioned as "built for agentic AI" — an accelerator designed for AI agent inference. Key specs: 350W TDP with air cooling, standard PCIe form factor, support for FP4 through FP64 precision. The reference design offers 160GB LPDDR5X, while partner-customized versions can go up to 480GB using 20 LPDDR5X chips across a 640-bit bus, delivering about 684 GB/s bandwidth at 10.7 Gbps. Sampling begins in the second half of 2026.
Everyone in the AI industry knows HBM is in critically short supply, but Intel chose a different path entirely. At Computex 2026, CEO Lip-Bu Tan personally unveiled Crescent Island — Intel's next-generation data center GPU. Its biggest selling point isn't raw compute or architecture, but 480GB of LPDDR5X memory. You read that right: LPDDR5X — the same memory chips found in laptops. While everyone else is fighting tooth and nail over HBM allocation, Intel is saying: skip HBM, use LPDDR5X, and pack in more capacity. It's a fascinating bet. Built on Intel's Xe3P GPU architecture, Crescent Island is officially positioned as "built for agentic AI" — an accelerator designed for AI agent inference. Key specs: 350W TDP with air cooling, standard PCIe form factor, support for FP4 through FP64 precision. The reference design offers 160GB LPDDR5X, while partner-customized versions can go up to 480GB using 20 LPDDR5X chips across a 640-bit bus, delivering about 684 GB/s bandwidth at 10.7 Gbps. Sampling begins in the second half of 2026.