Intel Crescent Island GPU Packs 480GB Without HBM: Xe3P Targets Inference Gap - Tech Times

At Computex 2026, Intel unveiled its Crescent Island data center GPU, a product that deliberately diverges from the prevailing AI hardware strategy by using LPDDR5X memory instead of high-bandwidth me... Read original →

Intel’s Crescent Island GPU, unveiled at COMPUTEX 2026, substitutes HBM with 480GB of LPDDR5X—a calculated maneuver around the HBM supply crunch exacerbated by geopolitical constraints and TSMC’s (Taiwan, China) CoWoS bottlenecks. Technically, this shifts data center GPU design toward standard PCIe 6.0 platforms, reducing reliance on advanced packaging and easing pressure on outsourced assembly capacity. However, its ~684 GB/s bandwidth caps it to lightweight inference, not large-model deployment. From a compliance standpoint, using mature-node LPDDR5X sidesteps emerging U.S.-South Korea export controls on HBM, enhancing supply chain security. NVIDIA will likely counter with aggressive H20 pricing to protect its inference stronghold. Over the next 18 months, if HBM3E scaling falters, LPDDR-based GPUs could become the fallback for second-tier players—but the training market remains irrevocably HBM-locked.