Skip to content
audiotale.net
Sign in
NewsNews

Intel GPU D-series: up to 480GB LPDDR5X, 1.5TB/s and 350W for AI inference, now sampling

Intel named its Crescent Island inference GPU at OCP 2026; the 480GB figure is a ceiling, and Intel's own card reportedly carries 160GB.

audiotale.net team

· 5 min read

Written by AI from the sources linked below. How we use AI

On this page
  1. Key points
  2. What happened
  3. Background
  4. The numbers
  5. What it means
  6. What we don't know yet
  7. What to watch next
  8. Sources

Intel has given its next data-centre inference GPU a name, the Intel GPU D-series (formerly code-named Crescent Island), and published its headline numbers: up to 480 GB of LPDDR5X memory, 1.5 TB/s of memory bandwidth and 1.3 petaflops of FP8 compute in a 350-watt, air-cooled PCIe card. Customer sampling is under way. The design is a deliberate bet that, for serving large models, how much memory fits on a card and what it costs to run matters more than peak bandwidth, and it is a different bet from the HBM-based accelerators that dominate the market.

Key points

  • Intel announced the GPU D-series at the Open Compute Project (OCP) Global Summit, in a Business Wire release dated 12 October 2026; some outlets date the coverage 13 October.
  • Intel's release lists up to 480 GB of LPDDR5X, 1.5 TB/s of bandwidth and 1.3 petaflops of FP8 compute at 350 W, on the Xe3P architecture.
  • StorageReview reports that the 480 GB figure applies only to partner-built cards and that Intel's own PCIe card carries 160 GB.
  • Intel says one card can run models of up to 300 billion parameters and an eight-GPU system more than one trillion.
  • Intel says customer sampling is under way; we found no price or general-availability date.

What happened

At OCP, Intel shared new technical details on the GPU D-series, introduced Xeon-based rack designs and described its open-software work. The company frames the chip around what it calls agentic AI: inference that is no longer a single prompt and answer but a multi-step workflow of retrieval, reasoning, orchestration and tool use. In that picture, power limits, memory bottlenecks and cost, rather than raw compute, decide how far AI can scale.

Intel's release describes the GPU as built on its Xe3P architecture and designed for "sustained, memory-intensive AI inference": large language models, long-context workloads and many concurrent inference jobs. It specifies a standard PCIe form factor with air cooling and the same "Xe software foundation" as other Intel GPUs, so that it fits existing data-centre environments without special racks or cooling. Anil Nanduri, an Intel vice president for AI products, is quoted as saying: "Inference is becoming the defining AI workload."

The catch is in the memory line. Intel's own release says "up to 480 GB". StorageReview's coverage of the announcement reports that the 480 GB capacity is available only on partner-built cards, and that Intel's own branded PCIe card has 160 GB. A second outlet, TechEnclave, reports the same split: 160 GB on Intel's standard card, 480 GB as a partner-built option. Intel's release does not make that distinction in the passage we read, so buyers should treat 480 GB as a ceiling and ask vendors which configuration they are quoted. Some headlines have run with the largest number alone.

Alongside the GPU, Intel said select Xeon 6+ configurations scale beyond 46,000 cores per rack at 133.45 kW, and listed day-0 software support for Meta's Muse Glimmer open model. StorageReview adds that the ecosystem slide named server makers including Dell, HPE, Lenovo and Supermicro and software partners including Red Hat, Canonical, vLLM and Hugging Face; that is the outlet's reading of the slide, not a list in the press text.

Background

The part was first shown as Crescent Island at OCP in 2025, and StorageReview notes the architecture was detailed again at Hot Chips this August, with 32 Xe cores and 256 XMX matrix engines. Its design choice, LPDDR5X instead of high-bandwidth memory (HBM), is the notable one. LPDDR5X is the low-power memory used in phones and laptops: it is cheaper and comes in large capacities, but it delivers far less bandwidth per device than HBM stacks.

That tradeoff matters because of how language models generate text. In the decode phase, each new token requires reading much of the model's weights from memory, so speed is often limited by bandwidth rather than arithmetic. Capacity matters for a different reason: the weights and the key-value cache for long contexts and many simultaneous users have to fit somewhere. A card with a lot of cheaper memory can hold a bigger model, or more sessions, at a lower cost per gigabyte, at the price of slower reads.

The numbers

Intel's comparison point in the coverage is its own Gaudi 3 PCIe card. StorageReview gives Gaudi 3 as 128 GB of HBM2e at 3.7 TB/s. Against that, the D-series trades bandwidth for capacity.

Memory capacity per card, GBGB
Memory capacity per card, GB (GB)
GB
Gaudi 3128
D-series, Intel card160
D-series, partner card480

Source: Intel press release (480 GB option); StorageReview (160 GB Intel card, Gaudi 3)

Memory bandwidth per card, TB/sTB/s
Memory bandwidth per card, TB/s (TB/s)
TB/s
Gaudi 33.7
GPU D-series1.5

Source: Intel press release (D-series); StorageReview (Gaudi 3)

Some arithmetic of our own, not Intel's, helps put this in context. At 1.5 TB/s, reading all 480 GB once would take about 0.32 seconds, and reading 160 GB about 0.11 seconds. A model that must stream most of its weights for every token could therefore see per-token latencies in that range on a single card, before any optimisation such as batching, quantisation or sparsity. Intel's 300-billion-parameter claim is consistent with the larger configuration: at one byte per parameter in FP8 the weights alone would be about 300 GB, which fits in 480 GB but not in 160 GB. This is our inference from the published figures; Intel has not said how it reached its model-size claim, and it has not published benchmark results that we could find.

What it means

For AI teams running inference, the appeal is cost and deployability: an air-cooled PCIe card that slots into ordinary servers avoids the liquid cooling and rack redesign that the highest-end accelerators increasingly require. The 350 W envelope is modest. For workloads where memory capacity decides what you can serve at all, such as large models with long contexts, more capacity per card could mean fewer cards.

For software, Intel's pitch rests on upstream support in frameworks and serving engines such as vLLM, and on code written for other Intel GPUs carrying over. How well that works in practice is the open question for any non-Nvidia part, and the answer will come from users, not from slides.

For the market, the announcement shows a split in accelerator design: HBM-heavy parts chase the lowest latency and the biggest training and serving clusters, while LPDDR-based parts target capacity per dollar and per watt. Which side of that line a workload falls on depends on its latency target and how many concurrent users it serves.

What we don't know yet

  • Price, general availability, and the volumes of the 160 GB versus 480 GB configurations.
  • Real-world throughput and latency: Intel has published specifications, not independent benchmarks.
  • Whether Intel's release or its partners will state clearly which cards carry 480 GB; we are relying on StorageReview and TechEnclave for the 160 GB figure.
  • How the software stack performs for popular models compared with Nvidia and AMD.

What to watch next

Watch for the first independent benchmarks from sampling customers, for OEM announcements naming 480 GB configurations, and for pricing. Intel has given no general-availability date that we could find.

Sources

SaveStart the discussion

Written by

audiotale.net team

The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.

Read next