Intel's long-lost data center prototype 'Arctic Sound' Xe-HP multi-tile GPU surfaces in new engineering sample — Company's cancelled AI processor features 32GB of HBM2E

Raja Koduri shows Intel Xe HP 4-tile GPU
The quad-tile variant of the Xe-HP GPU (Image credit: Intel (Hot Chips))

Back in 2020, Intel first showed off its multi-tile data center GPU codenamed "Arctic Sound," part of the Xe-HP (High Performance) family. It was supposed to represent one part of to be the company's grand entry into the AI GPU market. Arctic Sound was ultimately limited to internal samples because it was cancelled as a commercial product. Today, one of those original engineering samples seems to have landed in the hands of Chips By Layers on X, who's pictured it in all its glory.

Intel was working on a single-tile, dual-tile and a quad-tile variant of Arctic Sound; the picture above shows two tiles flanked by four memory modules on either side, which are 8GB HBM2E chips, so 32GB of VRAM in total. A single Xe-HP tile had 512 Execution Units (EUs) in its original config, but they were cut down for actual samples. This dual-tile part has 480 EUs per tile, or 960 EUs combined, which comes out to 7,860 shader cores. Arctic Sound 2T also had a 300W TGP.

We can clearly see "Intel Confidential" marked on the IHS, along with "QVS8 1.00 GHz." Funnily enough, the OP actually ordered a Ponte Vecchio GPU but received this engineering sample instead. Ponte Vecchio is another fascinating product in Intel's graphics history. It was part of the Xe-HPC (High Performance Computing) family and the true halo-tier AI GPU from the company, stitching together a 47-tile matrix featuring Xe cores instead of conventional Execution Units.

Latest Videos From

(Image credit: Intel)

Arctic Sound basically proved it was possible for EMIB to horizontally connect multiple GPU tiles, so Intel used that along with its Foveros 3D packaging tech to build Ponte Vecchio. Eventually, Arctic Sound was cancelled as a commercial product because it was simply too expensive to produce. The company reworked it into "Arctic Sound-M," which became the "Intel Data Center GPU Flex" series and it was moved to the Xe-HPG (High Performance Gaming) microarchitecture instead.

Next year, we expect to see Jaguar Shores, which can be looked at as a spiritual successor to both Ponte Vecchio and Arctic Sound since it scales the multi-tile concept to an unprecedented scale. Technically, though, it's more in-line with Ponte Vecchio since it's a flagship GPU, and that was already supposed to receive a direct successor in the form of Falcon Shores, but that lineup met the same fate as Arctic Sound last year.

The upcoming Cresent Island GPUs with LPRDR5X memory target a similar, cost-effective solution to data center compute that the original Arctic Sound tried to, so they're probably the closest in terms of a proper sequel.

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.

Hassam Nasir
Contributing Writer

Hassam Nasir is a die-hard hardware enthusiast with years of experience as a tech editor and writer, focusing on detailed CPU comparisons and general hardware news. When he’s not working, you’ll find him bending tubes for his ever-evolving custom water-loop gaming rig or benchmarking the latest CPUs and GPUs just for fun.

  • cp0x
    It's crazy that no one has bothered to release any products that can actually cost-effectively compete with Nvidia in this realm, despite Nvidia's ridiculously inflated prices. FFS, just put out a card with a decent vector processor, and 32GB, 48GB, 64GB, 96GB, 128GB, 256GB, and 512GB RAM. Nvidia is charging $13k for a 96GB card now. AMD or Intel could easily undercut Nvidia while still making bank. Maybe the AMD Instinct MI350P PCIe will be competitive, but unfortunately I think they'll charge way too much for it (for a card without Nvidia CUDA support).
    Reply
  • TerryLaze
    cp0x said:
    (for a card without Nvidia CUDA support).
    And there you go, you just answered your own question...the market for a card that doesn't support all the big nvidia stuff won't be big enough to justify the cost of research and development, let alone manufacturing.
    Reply
  • cp0x
    TerryLaze said:
    And there you go, you just answered your own question...the market for a card that doesn't support all the big nvidia stuff won't be big enough to justify the cost of research and development, let alone manufacturing.
    Unfortunately, it's a bit like the Windows monopoly in the 90s ... except that Windows was pretty bad at most things and Nvidia is pretty good at most things.
    Reply
  • bit_user
    cp0x said:
    just put out a card with a decent vector processor, and 32GB, 48GB, 64GB, 96GB, 128GB, 256GB, and 512GB RAM.
    Intel announced one that should support up to 480 GB.
    https://www.tomshardware.com/pc-components/gpus/intel-details-long-awaited-crescent-island-ai-gpu-at-computex-boasts-up-to-480-gb-of-lpddr5x-to-combat-memory-shortages-company-shares-more-details-of-its-xe3p-inference-accelerator-at-computex
    It's not as easy as you suggest. It requires the GPU be specifically designed to support LPDDR memory, and the reality is probably that a GPU chip can only support GDDR or LPDDR - not both. GDDR is clearly what you want for good gaming performance.

    cp0x said:
    Nvidia is charging $13k for a 96GB card now.
    The way they did it is by using a 512-bit memory interface and then putting 24-gigabit GDDR6 chips in a clamshell setup (i.e. sort of like double-rank). The top end GPUs from AMD and Intel (until the above Crescent Island launches) are 256-bit parts and will support a max of 48 GB.

    cp0x said:
    Maybe the AMD Instinct MI350P PCIe will be competitive,
    I doubt it. The die is huge and it uses HBM. It certainly won't be much cheaper than the 96 GB Nvidia card you mentioned, if at all.
    Reply