Intel Crescent Island GPU PCB Leak Reveals Xe3P AI Accelerator With 160GB LPDDR5X Memory
Intel’s upcoming Crescent Island PCIe accelerator has appeared in an early PCB leak, offering the first detailed look at what could become one of the company’s most interesting AI-focused data center products. Built around Intel’s next-generation Xe3P graphics architecture, Crescent Island is being designed for AI inference workloads, with a strong focus on performance per watt, lower deployment costs, and large memory capacity.
The leaked board appears to show a sizeable GPU package at the center, noticeably larger than Intel’s current Xe2-based BMG-G31 flagship graphics chip. The scale of the BGA pad suggests that Crescent Island is not a small client GPU repurposed for servers, but a dedicated accelerator aimed at enterprise AI workloads.
One of the most notable details is the memory configuration. Instead of using high-bandwidth HBM memory, Intel appears to be using LPDDR5X. The PCB reportedly includes 20 LPDDR5X memory package positions across the front and back of the board, with each module expected to offer 8GB. That would bring the total memory capacity to 160GB.
This is a major design choice. While rivals in the AI accelerator space are pushing expensive HBM3E and next-generation HBM solutions, Intel seems to be targeting a different balance of performance, capacity, power efficiency, and cost. LPDDR5X may not offer the same raw bandwidth as HBM, but it is easier to source, more affordable, and still capable enough for many AI inference workloads.
That strategy could make Crescent Island attractive for cloud providers, enterprise customers, and “tokens-as-a-service” businesses that need large-scale AI inference hardware without the extreme cost of top-tier training accelerators.
The leaked PCB also points to a high-end power delivery design. Around 13 VRM phases appear likely to be populated, while the board layout may support up to 18 in total. Power is supplied through a single 16-pin connector placed on the rear side of the board. A side-mounted USB Type-C port is also visible and is likely intended for engineering, debugging, or testing purposes rather than regular user functionality.
Crescent Island is based on Intel’s Xe3P architecture, an enhanced version of Xe3. Intel has already discussed Xe3 in relation to its future Panther Lake platform, but Xe3P is expected to scale beyond client graphics and into enterprise and data center AI hardware.
For consumer and client products, Xe3 is expected to appear in future Arc graphics offerings. Xe3P, however, appears to be the more flexible and powerful variant, designed to scale from integrated graphics all the way to dedicated data center accelerators.
Intel has positioned Crescent Island as a power-optimized and cost-optimized GPU for air-cooled enterprise servers. That matters because many data centers want AI accelerators that can be deployed without requiring exotic cooling or extremely expensive infrastructure upgrades. If Intel can deliver strong inference performance in a standard air-cooled server environment, Crescent Island could fill an important gap in the market.
The card is expected to support a broad range of data types, making it suitable for modern AI inference use cases. This includes workloads where large memory capacity, efficient throughput, and predictable operating costs are more important than peak training performance.
The decision to use 160GB of LPDDR5X memory may be one of Crescent Island’s biggest selling points. AI inference workloads, especially those involving large language models, can benefit heavily from more available memory. By offering a high-capacity memory pool without relying on costly HBM, Intel could make Crescent Island a practical alternative for companies focused on serving AI models at scale.
Intel is also working on its open and unified software stack for heterogeneous AI systems. Current Arc Pro B-series products are already being used to evaluate and improve these software foundations, which should benefit future accelerators like Crescent Island. Strong software support will be essential, as AI hardware success depends not only on silicon performance but also on developer tools, framework compatibility, driver maturity, and deployment simplicity.
Customer sampling for Intel Crescent Island is currently expected in the second half of 2026. That means more details should emerge over the coming months, including final specifications, performance targets, power limits, and the exact market positioning of the accelerator.
If the leaked PCB is close to the final design, Crescent Island could become a key part of Intel’s AI hardware roadmap. With Xe3P architecture, 160GB of LPDDR5X memory, PCIe connectivity, and a focus on cost-efficient inference, Intel appears to be preparing a product aimed at customers who want scalable AI acceleration without the premium pricing associated with HBM-based solutions.
