Intel Crescent Island leaked PCB reveals massive Xe3P GPU with 160GB LPDDR5X to dodge HBM crisis

Intel Crescent Island Takes Unconventional Memory Route

Intel Crescent Island, the company's next-generation data-center GPU, has appeared in its first leaked PCB images, revealing a bold architectural choice that sets it apart from competitors. The images, shared by @yuuki_ans on X, show the front and back of the PCB and confirm that Intel has opted for a single-GPU setup featuring the Xe3P GPU architecture paired with LPDDR5X memory rather than the high-bandwidth HBM standard used by rivals1

1

3

3

.

The leaked PCB reveals a massive GPU die that occupies a significant portion of the board, notably larger than Intel's current flagship Xe2-based BMG-G31. The chip takes up nearly the entire width of the PCIe x16 slot below it, signaling Intel's commitment to scaling its AI GPU capabilities2

2

. This marks a departure from dual-GPU configurations and represents a focused approach to delivering AI inference performance.

160GB LPDDR5X Memory Configuration Addresses HBM Shortage

The most striking feature visible in the leaked PCB is the memory configuration. Intel has positioned 20 LPDDR5X module sites across the board—12 on the front and 8 on the back—supporting a total of 160GB LPDDR5X memory using 8GB modules, the highest capacity currently available for LPDDR5X1

1

2

2

. This makes Intel Crescent Island the world's first AI GPU to use LPDDR5X, a strategic decision driven by the ongoing HBM shortage that has plagued the industry.

While competitors like the Nvidia H200 NVL with 141GB of HBM3 and AMD MI350P with 144GB of HBM3E offer substantially higher memory bandwidth—the H200 delivers nearly 5TB/s compared to Crescent Island's estimated sub-1TB/s with a 640-bit memory interface and 10.7Gbps memory speeds—Intel is betting on a cost-effective solution that addresses the supply constraints affecting HBM production1

1

3

3

. HBM memory sits at the center of the ongoing memory shortage crisis, as it remains the hardest to secure for GPU manufacturers. By choosing LPDDR5X, Intel sidesteps these supply chain bottlenecks while keeping production costs down.

Power Delivery and Design Details for Air-Cooled Servers

The leaked PCB also reveals a robust power delivery system designed to support the massive Xe3P GPU. The board features a single 16-pin 12V-2x6 power connector positioned at the rear, along with 18 VRM positions, of which 13 appear populated in the current design2

2

3

3

. A USB Type-C port visible on the side appears intended for testing purposes, suggesting this PCB is close to its final production form.

Intel has positioned this AI GPU specifically for air-cooled servers and workstations, targeting AI inference workloads rather than training tasks. The Xe3P architecture powering Crescent Island is designed for power efficiency and cost optimization, supporting a broad range of data types ideal for Tokens-as-a-service providers and inference use cases at scale3

3

. This focus on inference accelerator capabilities positions Intel to compete in a rapidly growing segment where lower memory bandwidth may be acceptable if offset by capacity, cost advantages, and power efficiency.

Market Positioning and Timeline

Intel announced Crescent Island in October 2025, confirming it would feature the Xe3P graphics architecture following the current Xe2 architecture. Notably, Xe3P will not appear in consumer Celestial gaming graphics cards, as Intel canceled that product line. Instead, the architecture will scale from client integrated GPUs to data center AI GPUs, with Crescent Island representing the high-end implementation2

2

.

Intel plans to start customer sampling in the second half of 2026, giving the company time to refine the design and optimize its software stack1

1

2

2

. The company is already evaluating its open and unified software stack for heterogeneous AI systems with its existing Arc Pro B-series lineup, which should enable smoother deployment for early adopters.

The choice to use LPDDR5X memory represents a calculated trade-off. While memory bandwidth remains one of the most important factors in AI GPU performance for machine learning workloads, Intel appears confident that its architecture optimizations and cost advantages will appeal to customers focused on inference rather than training. As NVIDIA and AMD continue pushing HBM3E and teasing HBM4 for upcoming chips like Rubin and MI400, Intel's alternative approach could carve out a distinct market position—provided the Xe3P GPU delivers sufficient performance despite the bandwidth limitations. Watch for performance benchmarks and pricing announcements as customer sampling approaches later this year.