Intel leans on LPDDR5X to dodge global HBM crisis, leaked Crescent Island AI GPU pics reveal massive Xe3P core — chip sidesteps HBM shortage with 160GB of cheaper memory

News
By published

Intel is going with one massive GPU for its next-generation Xe3P AI GPU, and LPDDR5X memory rather than HBM.

Intel Headquarters, with people walking by
(Image credit: Getty Images / Justin Sullivan)

Pictures of the PCB of Intel’s upcoming Crescent Island data center GPU, announced last year, have surfaced. The images, which come by way of @yuuki_ans on X, show the front and back of the Crescent Island PCB and reveal the layout of the board's components, including the GPU, VRAM, and power delivery system.

The PCB shots confirm that Intel has opted to go with a single-GPU setup for Crescent Island, rather than a dual-GPU setup. A vast majority of the PCB’s real estate is consumed by a massive GPU socket in the middle, taking up the width of almost the entire PCIe x16 slot below.

There are pads for twelve LPDDR5X modules to the top, left, and right of the GPU socket and an additional eight on the back of the PCB, totaling 20. This confirms the GPU will use 32GB (8 GB) modules, the highest capacity modules out there for LPDDR5X currently on the market. The power delivery system consists of a single 16-pin 12V-2x6 power connector on the right and 19 power phases.

Latest Videos From

Crescent Island is the codename for Intel’s next-generation data center GPU powered by its Xe3P GPU architecture. Its most noticeable trait is the inclusion of LPDDR5X memory over HBM, which should make Crescent Island the world’s first AI GPU to use LPDDR5X. This will give Crescent Island significantly inferior memory bandwidth, but undoubtedly, Intel is opting to use slower memory to save on production costs, thanks to the memory shortage. (HBM memory sits at the center of the outgoing memory shortage crisis, as it's the hardest to secure for GPU manufacturers .)

Assuming the GPU comes with a 640-bit memory interface and uses 10.7Gbps memory, maximum memory bandwidth will be well under 1TB/s — a far cry from Nvidia’s older H200 GPUs, which have nearly 5TB/s of memory bandwidth. Memory bandwidth is one of the most important aspects of an AI GPU's performance and how quickly it can execute machine learning workloads, so it will be interesting to see how much LPDDR5X impacts Crescent Island’s performance compared to HBM-equipped cards.

Intel’s new AI-GPU will be targeted at air-cooled servers and will be a competitor to existing Nvidia and AMD cards in the same target demographic, such as the recently released AMD MI350P with 144GB of HBM3E and the Nvidia H200 NVL with 141GB of HBM3. Intel plans to start sampling Crescent Island to its customers in the second half of 2026.

Google Preferred Source

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.

Aaron Klotz
Aaron Klotz
Contributing Writer

Aaron Klotz is a contributing writer for Tom’s Hardware, covering news related to computer hardware such as CPUs, and graphics cards.

5 Comments Comment from the forums
  • usertests
    This confirms the GPU will use 32GB (8 GB) modules, the highest capacity modules out there for LPDDR5X currently on the market.
    32 GB/Gb != 8 GB, and it's wrong anyway, since we've just seen that Gorgon Halo is using 24 GB LPDDR5X modules, 8 of them, to reach 192 GB.

    There's 20 modules on the card. Leakers are saying Crescent Island can also use 24 GB modules, to triple the current capacity to 480 GB. Or 16 GB modules for 320 GB.

    https://videocardz.com/newz/intel-crescent-island-gpu-to-support-lpddr5x-9600-memory-and-1-5-tb-s-bandwidth2057321586642858006View: https://x.com/yuuki_ans/status/2057321586642858006
    2057352001344983485View: https://x.com/harukaze5719/status/2057352001344983485
    Reply
  • Notton
    It's basically the same idea as Apple's M Max/Pro and AMD's next gen gaming GPUs, and Strix/Medusa Halo.

    What remains to be seen is LPDDR5X availability by the time the new chips get released.
    Currently, it's bad enough that Samsung reduced production volume of their Galaxy smartphones just to meet data center demands.
    Reply
  • thestryker
    With an add in card like this there's nothing particularly stopping them from using even high density LPDDR5X. My assumption is that they will be doing capacity figures based on customer feedback. For a fully populated board the lowest capacity possible is 80GB and highest is 1280GB so there's plenty of choices. My guess is that trying to get high density is a losing proposition due to its use with SOCAMM modules so anything above 480GB seems unlikely.
    Reply
  • usertests
    thestryker said:
    With an add in card like this there's nothing particularly stopping them from using even high density LPDDR5X. My assumption is that they will be doing capacity figures based on customer feedback. For a fully populated board the lowest capacity possible is 80GB and highest is 1280GB so there's plenty of choices. My guess is that trying to get high density is a losing proposition due to its use with SOCAMM modules so anything above 480GB seems unlikely.
    I don't think anyone has come up with the Xe3P core counts for this, but it seems like it should be a lot faster than Strix/Gorgon Halo. All I see is some speculation on AnandTech Forums that it has 32 compute blocks, 384 Xe cores.

    Gorgon Halo is allowing the allocation of 160 GB VRAM out of 192 GB unified memory, the exact amount of the initial Crescent Island product. So there could be a strong case to go further to 240 GB (20x12) or 320 GB (20x16). It depends on what size models customers want to run on it.

    Are you sure about SOCAMM? The size looks about right but I don't see a middle M2 screw.

    https://pbs.twimg.com/media/HIsbeDZW8AEtPtb?format=jpg&name=large
    https://www.servethehome.com/wp-content/uploads/2025/03/NVIDIA-Grace-with-Micron-SOCAMM-Memory-without-Package.jpg
    Reply
  • thestryker
    usertests said:
    Are you sure about SOCAMM? The size looks about right but I don't see a middle M2 screw.
    I am referring to high density LPDDR5X packages only being used in SOCAMM. To my knowledge packages above 24GB have not appeared anywhere else, but SOCAMM goes up to 64GB.
    usertests said:
    I don't think anyone has come up with the Xe3P core counts for this, but it seems like it should be a lot faster than Strix/Gorgon Halo. All I see is some speculation on AnandTech Forums that it has 32 compute blocks, 384 Xe cores.
    RZL-AX is supposed to be in 16/32 Xe3P core configs so I'd think 32 would be minimum for these cards. I know Xe3P should be a lot better than Xe2, but for a high capacity enterprise card I'd think maybe 48-64 could be on the table.
    Reply