Intel officially launched the GPU D-Series on October 12, introducing the GPU D270 as the first product in a new line of inference-optimized data center GPUs. According to Intel's product page, the D270 is built specifically for agentic AI workloads, long-context processing, and high-concurrency serving "at practical power and cost" — and it arrives with a headline specification that no rival currently matches: up to 480GB of LPDDR5X memory on a single PCIe card.
The launch marks the transition of the Crescent Island accelerator from a technical blueprint into a shippable product, and it signals that Intel intends to compete in the fastest-growing segment of the AI hardware market rather than fight Nvidia on its home turf of high-end training silicon. You can track how the broader race for AI data center hardware unfolds in our ongoing AI news coverage.
From Hot Chips Preview to Official Product
Intel first detailed the Crescent Island design at Hot Chips 2026 in August, where it pitched a 480GB AI inference GPU aimed squarely at the memory-hungry demands of modern inference. What changed this week is the formal productization: Intel has now published the D-Series product line with the D270 as its inaugural SKU, and the launch was quickly picked up by Tom's Hardware, Phoronix, VideoCardz, StorageReview, and Wccftech.
Tom's Hardware reported that the chip delivers 1.3 peak PFLOPS of FP8 compute and 1.5 TB/s of memory bandwidth, positioning the D270 as the commercial incarnation of the Crescent Island accelerator. VideoCardz confirmed the part is based on Intel's Xe3p architecture — the same IP family that underpins the company's latest Arc graphics — while Notebookcheck noted that the card deliberately skips 3D graphics output entirely. This is a compute device for data centers, full stop.
The Specifications
The numbers Intel is publishing for the GPU D270 paint a clear picture of its intended workload:
- Compute: 1.3 peak PFLOPS of FP8, the precision that dominates modern LLM inference
- Memory: up to 480GB of LPDDR5X — several times the capacity of a typical HBM-equipped accelerator card
- Memory bandwidth: 1.5 TB/s
- Power: approximately 350W on a standard PCIe card
- Architecture: Intel Xe3p, the "Crescent Island" design unveiled at Hot Chips 2026
Wccftech characterized the design as "value-oriented," emphasizing how Intel has squeezed 1.3 PFLOPS of FP8 into a 350W envelope, while StorageReview highlighted the combination of 480GB LPDDR5X and high FP8 throughput in a standard PCIe form factor. The card does not require exotic cooling or proprietary interconnects, which lowers the barrier for cloud providers and enterprises to adopt it in existing servers.
Why LPDDR5X Instead of HBM
The most consequential engineering decision in the D270 is the use of LPDDR5X rather than HBM (high bandwidth memory). HBM offers far greater bandwidth per module, but it is expensive, supply-constrained, and capacity-limited — typically capping accelerator cards at 80GB to 192GB. LPDDR5X is cheaper, far more plentiful, and allows dramatically higher total capacity, at the cost of lower peak bandwidth.
Intel's argument, laid out on the product page, is that this trade-off matches how modern AI workloads actually behave: "Agents read far more than they write: long contexts, multistep workflows, compute dominated by prefill." In other words, agentic AI is less about the raw speed of generating tokens and more about ingesting enormous contexts — codebases, documents, tool outputs — quickly enough to keep models fed. Intel's tagline for the line is "Scale inference, not infrastructure."
For batched, high-concurrency serving — where many user requests are processed in parallel — memory capacity becomes the constraint that determines how many simultaneous contexts fit on one card. A 480GB card can hold far more concurrent state than a 141GB or 192GB competitor, which can translate into better utilization per dollar even if peak bandwidth is lower.
An Inference-First Challenge to Nvidia
Yahoo Finance framed the launch as Intel aiming "to boost AI turnaround with new GPU chips to take on Nvidia" — and the framing is apt. Nvidia's data center dominance rests on high-margin, HBM-laden accelerators coupled with its CUDA software ecosystem. Rather than competing head-on for the same training budgets, Intel is attacking the inference tier of the market, where cost per token and memory capacity per watt increasingly decide purchasing decisions.
The strategy mirrors moves by other chipmakers who have concluded that the agentic AI era changes the hardware calculus: if agents spend most of their compute reading rather than writing, then capacity and cost efficiency matter as much as raw FLOPS. The D-Series branding also implies a family of products rather than a one-off — Intel's product page is structured as a line introduction, with the D270 as the first entry.
What It Means for AI Builders
For teams running inference at scale, the D270's pitch is straightforward: more memory per card, reasonable power draw, and standard PCIe integration mean potentially lower cost per concurrent user for long-context and agentic applications. RAG pipelines, coding agents that carry entire repositories in context, and document-heavy enterprise workflows are the obvious beneficiaries.
The open questions are the ones that always accompany a new accelerator launch: Intel has not yet published pricing, broad availability dates, or a software stack benchmark — and in AI hardware, the software ecosystem is frequently the deciding factor. How the D270 performs in real inference stacks, against Nvidia's latest parts and against rising Chinese accelerators in their home market, will determine whether Crescent Island is a niche curiosity or the beginning of a genuine third option in AI data center silicon.
