AnswersPublished20 sources
Intel’s GPU D-Series Brings Crescent Island to AI Inference
At OCP on October 12, Intel named its Crescent Island accelerator the GPU D Series and announced up to 1.3 PFLOPS of FP8 compute, 480GB of LPDDR5X memory and 1.5TB/s bandwidth. The 350W PCIe card is designed for air cooling and AI inference; reported memory configurations range from 160GB on Intel branded versions t...
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- At OCP on October 12, Intel named its Crescent Island accelerator the GPU D Series and announced up to 1.3 PFLOPS of FP8 compute, 480GB of LPDDR5X memory and 1.5TB/s bandwidth.
- The 350W PCIe card is designed for air cooling and AI inference; reported memory configurations range from 160GB on Intel branded versions to as much as 480GB on partner cards.[6][11]
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did Intel announce at the October 2026 OCP Global Summit in San Jose about its Crescent Island accelerator, officially named the Intel. Article summary: Intel’s October 12, 2026 OCP announcement gave Crescent Island its commercial name, Intel GPU D-series, and disclosed an inference-focused Xe3P accelerator targeting 1.3 petaflops of FP8 compute with up to 480GB of LPDDR. Topic tags: general, general web, user generated, news. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts wi
Intel gave its Crescent Island accelerator a product name at the October 12, 2026, OCP Global Summit: Intel GPU D-Series. Built on the Xe3P architecture, the accelerator is designed for sustained, memory-intensive AI inference, with Intel advertising up to 1.3 petaflops of FP8 compute, 1.5TB/s of memory bandwidth and up to 480GB of LPDDR5X memory.1
The pitch is as much about deployment as peak compute. The D-Series is described as a 350W PCIe accelerator with air cooling, while the available sources do not provide independent workload benchmarks to show how it performs in practice.9
11
Intel GPU D-Series specifications
| Specification | Announced or reported detail |
|---|---|
| Architecture | Intel Xe3P |
| Intended workloads | AI inference, including sustained, memory-intensive workloads |
| FP8 compute | Up to 1.3 petaflops; an advertised peak figure, not a measured inference result |
| Memory | Up to 480GB LPDDR5X |
| Memory bandwidth | 1.5TB/s |
| Power and form factor | 350W PCIe accelerator, described as air-cooled |
| Sampling | Customer sampling was reported underway in coverage of the announcement |
Memory capacity varies by card
The headline 480GB figure is a maximum, not a guarantee that every D-Series card will ship with that capacity. Earlier reporting describes Intel-branded cards with 160GB of LPDDR5X and partner-built versions with up to 480GB.6
11 That distinction matters for buyers evaluating how much model data a card can hold, but memory capacity alone does not establish serving speed or concurrency.
What the architecture reports add
Technical coverage published before the OCP announcement described Crescent Island as having 32 Xe3P cores, 256 XMX matrix engines and 32MB of L2 cache.5
8
9 Those details help explain the accelerator’s compute design, but Intel’s headline OCP figures remain the clearest announced measures for comparing its memory capacity, bandwidth and FP8 peak.
Why Intel emphasizes air cooling
A 350W air-cooled PCIe card gives Intel a deployment-focused story: the accelerator is presented in a conventional add-in-card form rather than as a component that requires a liquid-cooled system.9
11 That may make it relevant to organizations considering inference capacity within existing data-center designs. It does not, by itself, establish lower total ownership costs; that would also depend on purchase price, system requirements, software support and real workload performance.
Intel’s announcement is not enough to make a like-for-like performance verdict against Nvidia’s Rubin products. The available material does not provide independent benchmarks or comparable measured inference results. Peak FP8 figures should not be treated as a forecast of model-serving speed, latency or cost per token.
Model capacity claims need context
Coverage of Intel’s announcement reports support for models with up to 300 billion parameters on one card and more than one trillion parameters in an eight-GPU configuration.2 These are model-capacity claims, not performance guarantees. The sources provided do not specify enough about precision, runtime memory use or workload conditions to translate them into a dependable estimate of throughput or latency.
Intel also announced Xeon rack designs
Alongside the GPU details, Intel introduced new Xeon-based rack designs developed with ecosystem partners.1
17 Reporting on selected Xeon 6+ configurations cited more than 46,000 CPU cores per rack and 133.45kW of total rack power; that power figure refers to the rack, not an individual processor.
20
Intel’s broader message, as summarized in its announcement, was that enterprises need to match workloads with suitable compute and infrastructure. The D-Series is one part of that strategy: an inference-focused accelerator alongside Xeon systems, rather than a claim that a single card suits every AI workload.17
What the announcement leaves open
The cited materials do not establish pricing, a confirmed general-availability date or independent performance results. While customer sampling was reported underway, buyers will need further information to assess availability, software compatibility and performance on their own inference workloads.7
For now, the clearest takeaway is a product proposition, not a proven cost or speed advantage: Intel is targeting AI inference with a high-memory, 350W air-cooled PCIe accelerator. How compelling it is will depend on shipping configurations and measured results—not headline specifications alone.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Intel’s GPU D-Series Brings Crescent Island to AI Inference"?
At OCP on October 12, Intel named its Crescent Island accelerator the GPU D Series and announced up to 1.3 PFLOPS of FP8 compute, 480GB of LPDDR5X memory and 1.5TB/s bandwidth.
What are the key points to validate first?
At OCP on October 12, Intel named its Crescent Island accelerator the GPU D Series and announced up to 1.3 PFLOPS of FP8 compute, 480GB of LPDDR5X memory and 1.5TB/s bandwidth. The 350W PCIe card is designed for air cooling and AI inference; reported memory configurations range from 160GB on Intel branded versions to as much as 480GB on partner cards.[6][11]
Sources
intel.comIntel Accelerates Open Infrastructure for the Agentic AI Era
videocardz.comIntel details Xe3P GPU architecture, Crescent Island gets up to 480GB memory and 350W PCIe variant - VideoCardz.com
storagereview.comNews Archive
fourweekmba.comIntel Details Crescent Island GPU and New Xeon Rack Designs - FourWeekMBA