Quick answer: Intel's new GPU D-series (codename Crescent Island) is an AI inference card for servers, not a gaming GPU. Its first model, the D270, has 160GB of LPDDR5X on Intel's own card, and Intel officially allows partner cards with up to 480GB. A leaker now claims partners are building 512GB boards, but Intel has not confirmed this. There is no price or India launch date yet.
Label guide: OFFICIAL = stated by Intel · REPORTED = from press reports, not an Intel spec · RUMOUR = unconfirmed leak · TECHCURE RECOMMENDATION = our advice. Last checked: 17 October 2026.
What is the Intel GPU D-series?
OFFICIAL Intel announced the GPU D-series at the OCP Global Summit on 12 October 2026. It is the product name for the accelerator previously called Crescent Island, and the first card is the Intel GPU D270. It is built on Xe3P, an updated version of Intel's Xe3 graphics architecture, and is designed for running AI models (inference) such as large language models, long-context tasks and many AI requests at the same time.
Intel removed the dedicated 3D graphics and ray-tracing hardware to make room for more AI compute. That is why this card will not run games like an Arc or GeForce card.
Intel GPU D270 specifications
| Specification | Intel GPU D270 | Label |
|---|---|---|
| Architecture | Xe3P, 32 Xe3P cores, 256 XMX matrix engines | OFFICIAL |
| AI compute | Up to 1.3 PFLOPS FP8 (theoretical) | OFFICIAL |
| Memory bandwidth | 1.5 TB/s | OFFICIAL |
| Memory – Intel card | 160GB LPDDR5X | OFFICIAL |
| Memory – partner (ODM) cards | Up to 480GB LPDDR5X | OFFICIAL |
| 512GB partner cards | Claimed by a leaker, not confirmed | RUMOUR |
| Power and cooling | 350W, air-cooled (passive) for standard servers | OFFICIAL |
| Interface | PCIe 5.0 add-in card | OFFICIAL |
| Price / availability | Customer sampling started; price and general availability not announced | OFFICIAL |
Why so much memory? LPDDR5X instead of HBM
Most data-centre AI accelerators use HBM memory, which is very fast but expensive and in short supply. Intel chose LPDDR5X, the same family of memory used in laptops and phones. It is slower per chip than HBM, but it lets one card carry far more capacity at lower power and cost.
- REPORTED Leaked board photos showed 20 memory package positions. With 8GB packages that gives Intel's 160GB; with 24GB packages it gives the official 480GB maximum.
- RUMOUR On 13 October 2026 the leaker Haze (@Haze2K1) posted that 512GB partner boards also exist. No manufacturer or specification was named, and reaching 512GB would need a different memory layout that Intel has not described.
What can 480GB actually run?
OFFICIAL Intel says one card can serve language models of roughly 300 billion parameters, and an eight-GPU server can run models beyond 1 trillion parameters.
ESTIMATED Simple maths shows why: in FP8, each parameter needs about 1 byte, so a 300-billion-parameter model needs roughly 300GB just for its weights, plus extra memory for long conversations (the KV cache). For comparison, NVIDIA's RTX PRO 6000 Blackwell – one of the largest workstation cards sold today – has 96GB.
Is it a gaming or workstation card?
No for gaming. It sits between a server accelerator and a workstation card: it is a standard PCIe card, and Intel plans partner-designed versions, which is unusual for server GPUs. But it targets AI servers. Intel also showed other form factors at OCP, including a liquid-cooled module with two GPUs REPORTED. The Xe3P architecture itself is planned for future Intel PC graphics too, but those will be different products.
What this means for AI PC buyers in India
TECHCURE RECOMMENDATION
- Don't wait for it for a personal AI PC. There is no price, no India availability and no retail channel announced. It is built for data-centre servers.
- For local LLMs and Stable Diffusion today: GPU memory (VRAM) still decides which models you can run. Compare options in our AI PCs & Deep Learning Workstations and NVIDIA RTX PRO workstation cards, or read the VRAM guide for local AI.
- Businesses planning on-premise AI servers: keep the D-series on your shortlist and wait for real pricing and independent benchmarks. We will update this post when Intel shares India details.
- Prefer Intel graphics for a desktop today? See Intel Arc graphics cards, or plan a full build on our AI PC for Stable Diffusion & Local LLMs page.
Planning an AI workstation or local LLM server? TechCure India builds custom AI and deep-learning workstations with pan-India delivery. Tell us the models you want to run and your budget – we will suggest the right GPU and memory.
FAQ: Intel GPU D-series
What is Intel Crescent Island?
Crescent Island was the codename for Intel's AI inference accelerator. It is now sold as the Intel GPU D-series, and the first model is the GPU D270.
How much memory does the Intel GPU D270 have?
Intel's own card has 160GB of LPDDR5X. Intel allows partner (ODM) cards with up to 480GB.
Is there a 512GB Intel GPU D-series card?
Not officially. A leaker claimed on 13 October 2026 that partners are building 512GB boards, but Intel has not confirmed it and no manufacturer was named.
Can the Intel GPU D270 play games?
No. Intel removed the dedicated 3D graphics and ray-tracing hardware. It is designed only for AI inference in servers.
When will the Intel GPU D-series launch in India and what will it cost?
Intel has started customer sampling but has not announced pricing or general availability, including for India.
Sources
- VideoCardz – 512GB partner boards reportedly planned
- Phoronix – Intel announces the GPU D-series
- ServeTheHome – GPU D-series name and specs
- Notebookcheck – D270 skips 3D graphics
- Guru3D – D270 model-size claims
Disclaimer: Intel, Arc and Xe are trademarks of Intel Corporation; NVIDIA and RTX are trademarks of NVIDIA Corporation. TechCure India is an independent PC builder and retailer and is not affiliated with or endorsed by Intel or NVIDIA. TechCure India does not currently sell the Intel GPU D-series. Information is based on public announcements and reports as of 17 October 2026 and may change. Featured image: original illustration by TechCure India of a generic PCIe AI accelerator card – not a photo of the actual Intel product.