Monday, August 24th 2026
Intel Details "Crescent Island" Graphics: 32 Xe3P Cores, up to 480 GB LPDDR5X Memory
Intel's long-awaited graphics card update is arriving for data centers and workstations with the introduction of "Crescent Island," now detailed. At Hot Chips 2026, the company revealed more about the upcoming graphics card. Based on the Xe3P GPU IP, the graphics card is designed as a 350 W, air-cooled, PCIe format compatible with any configuration. The GPU itself features 32 Xe3P cores, 256 XMX engines, 32 MB of unified L2 cache, and 160 GB of LPDDR5X memory. This memory configuration is exclusive to Intel-branded "Crescent Island" graphics cards, while the company allows its ODMs to equip the card with up to 480 GB of LPDDR5X memory for maximum VRAM capacity. Just a few of these in their maximum configuration could power frontier open-source models locally.
What is particularly interesting about Xe3P is the pipeline, as XMX engines have been redesigned from the four-deep systolic design on the regular Xe3 to 16-deep processing on Xe3P. Additional details include FP4 data format support alongside FP8, as well as massive register files of 1 MB per core and L1 cache of 512 KB per core. In total, there are 256 Vector engines and 256 of the new XMX engines, each split into groups of eight per Xe3P core. Additionally, in Xe3P, there are no graphics processing elements that would output any imagery, as the silicon is designed purely for AI inference workloads that include long context, agentic tasks.
For a GPU with the power of "Crescent Island," Intel targets users in PC, workstation, and data center edge deployments, where fast AI inference is essential. For example, running a trillion-parameter model, like Kimi K3, on local hardware would require nearly 2 TB of aggregate system memory. With four GPUs each carrying 480 GB, that goal would be achievable on a single workstation. Of course, that memory configuration wouldn't come cheap, as DRAM is very expensive. However, it remains a possibility, especially since Intel is offering day-zero support for all major frameworks, and software integration is already in place.
What is particularly interesting about Xe3P is the pipeline, as XMX engines have been redesigned from the four-deep systolic design on the regular Xe3 to 16-deep processing on Xe3P. Additional details include FP4 data format support alongside FP8, as well as massive register files of 1 MB per core and L1 cache of 512 KB per core. In total, there are 256 Vector engines and 256 of the new XMX engines, each split into groups of eight per Xe3P core. Additionally, in Xe3P, there are no graphics processing elements that would output any imagery, as the silicon is designed purely for AI inference workloads that include long context, agentic tasks.
18 Comments on Intel Details "Crescent Island" Graphics: 32 Xe3P Cores, up to 480 GB LPDDR5X Memory
Also, "GPGPU" kinda just disappeared in use despite GPUs being most-commonly used today for AI/MI 'General Purpose' compute. So, maybe they'd say "Generative Processing Unit"?
Well, I'll wait for Asrock to throws us a LPDDR3 X Spaced, 12 GB @96 bit next to NPU, so we can afford it while JH will just sold another jacket to Lisa.
That means a relatively slow inference speed.
Not saying there's a conspiracy here, just curious why it was even included.
Two of these would allow you to run very decent models, hell, even one of these could run a smaller quant of DeepSeek V4 Flash comfortably.
As much as 395+ MAX APU has the 8060S equivalent to mobile 4060 and desktop 3060. Yet they advertised it as equivalent with mobile 5060 . LPDDRX on MAX is 8000 quad channel, I personally didn't see it dropping under but maybe it does under certain stress scenarios.I think is only 128 bit on P Lake.
In fact you could run this in 1-bit quantization on one of these...barely
Novalake is using the Xe3p cores... so there are definitely Xe3p cores with display out.