TL;DR · 30-second read
The Short Version
Intel has shown a new computer chip for running AI services such as chatbots. It holds more working memory than the storage drive in a typical laptop, and about two-thirds more than the Nvidia chip that previously led on memory.
Surprisingly, it uses the type of memory found in laptops. It is also cooled by air, so it could fit inside many existing computer buildings.
That means enormous AI programs can fit on one card. Whether they will respond quickly enough is unproven, and Intel has not announced a price or release date.
TechRadar reported on October 18 that Intel unveiled the Intel GPU D-series, previously codenamed Crescent Island, at the recent Open Compute Project Global Summit. The AI inference accelerator offers up to 480GB of LPDDR5X memory, ahead of the 288GB of HBM3e on Nvidia’s Blackwell Ultra B300, which the report identified as the previous capacity leader among data center GPUs.
Intel says one card can support large language models of up to 300 billion parameters, while an eight-GPU system can accommodate models exceeding one trillion. The company also specifies 1.3 petaflops of FP8 compute, 1.5TB/s of memory bandwidth, a 350W power envelope, a standard PCIe form factor and air cooling. It has not announced a public price or a specific general-availability date.
Executive Summary
Intel’s GPU D-series is built for inference, the stage at which a trained AI model answers requests. Rather than using the GDDR7 memory associated with modern graphics cards or the HBM3e used on Nvidia’s B300, Intel chose LPDDR5X, a power-conscious memory type commonly found in laptops. The result is 67% more capacity than the B300’s 288GB, in a card designed to fit existing data center infrastructure.
The announcement matters because it presents a distinct theory of how to compete in a market dominated by Nvidia: make very large models easier to host, keep the power and cooling requirements modest, and sell on deployment economics. Intel’s Anil Nanduri, vice president of AI Products and GTM in the Data Center Group, described the goal as “making high-capacity inference practical to deploy and economical to scale.”
The specifications establish that large models should fit. They do not establish how quickly those models will respond or how many users each card can serve. Those results, together with the price Intel has yet to announce, will determine whether the capacity advantage translates into an economic one.
Capacity Decides What Fits. Bandwidth Shapes How Fast It Runs.
The most useful way to read the D-series is to set its 480GB of capacity against its 1.5TB/s of memory bandwidth, the rate at which data can move between memory and the processor. A model’s parameters are the numerical values it learned in training. At FP8, an 8-bit number format that uses one byte per value, a 300-billion-parameter model would occupy about 300GB. That would leave roughly 180GB for the working data that accumulates around long prompts and simultaneous requests. Intel has not said what precision its 300-billion figure assumes, so this is an illustration rather than a specification.
The same arithmetic shows the constraint. Moving 300GB at 1.5TB/s takes about 0.2 seconds, implying a rough ceiling of five full passes through those weights per second. In a dense model, where generating each token (a word fragment) touches most of the weights, that ceiling bears directly on response speed. Batching several requests into a single pass can raise total output, while model designs that activate only part of their weights can reduce the data moved per token. Cutting the model to 50GB would raise the same theoretical ceiling to about 30 passes per second.
The operational conclusion is that buyers should size this card on measured throughput for their own models, not on the largest model it can hold. A very large model that fits may suit long-context or batch-oriented work while proving less suitable for tasks that need immediate replies. That matters for Intel’s own target list, which includes both lengthy prompts and robotics applications requiring responsive processing. The 1.3-petaflop rating, or 1.3 quadrillion FP8 operations per second, is also a peak hardware figure rather than a measure of what a particular model will deliver.
Laptop Memory Is a Deliberate Trade-Off
Using LPDDR5X in a data center accelerator looks unusual only if the card is judged as a conventional graphics product. The choice suggests Intel is prioritizing capacity and power efficiency for inference over the traits expected of a gaming GPU. It also separates the D-series from the B300, whose HBM3e is a stacked memory type designed for high bandwidth.
That difference limits what the headline comparison can show. A 480GB-to-288GB gap is a clear capacity lead, but it is not a performance ranking. Comparing the two products fairly would require testing the same models, at the same precision and response targets, while measuring output per card, per watt and per dollar.
The 350W, Air-Cooled Design Is Aimed at Existing Halls
The less eye-catching specifications may matter most to facility operators. A 350W power envelope, standard PCIe card format and air cooling are intended to let the D-series enter existing data centers without specialized cooling arrangements or entirely new rack infrastructure. For an enterprise or colocation provider without liquid cooling, that could widen the set of sites where large-model inference is feasible.
Scale still needs planning. Eight cards would draw up to 2.8kW for the GPUs alone, before counting processors, networking, storage and fans. At maximum memory, the same system would hold 3.84TB. Whether that fits comfortably depends on each facility’s power and airflow limits per rack, and on how many such systems an operator wants to place side by side.
Can Capacity Win Share From Nvidia?
Intel’s slogan, “put every prompt in its place,” describes a routing strategy more than a head-on performance contest. Nicolas Dubé, corporate vice president of Intel’s Data Center Systems and Solutions Group, said organizations need infrastructure that can match workloads with “the right intelligence, right compute resources and right economics.” Intel also plans to pair the accelerator with Xeon processors, which would coordinate tasks, manage data movement and handle work that does not need a GPU, an approach aimed at AI systems built around retrieval, reasoning, orchestration, external tools and execution.
That gives the D-series a plausible opening with enterprises running open models, particularly where memory capacity and facility limits are the binding constraints. It does not need to beat Nvidia on every workload to find a role. It does need to show that its total cost per unit of useful output is competitive, and that moving workloads onto it is practical. Intel’s promise of software compatibility across its own GPU platforms addresses part of that challenge, but adoption will depend on how readily customers’ existing models and tools run in production.
Background
GPUs, originally developed to render graphics, are now widely used for AI because they can perform many calculations at once. As deployed models grow and prompts lengthen, memory has become a major design consideration: a model’s weights and associated data need to remain accessible while it serves requests. Accelerator makers use different memory technologies to balance capacity, bandwidth and power, including HBM3e on Nvidia’s B300, GDDR7 on many modern graphics cards and LPDDR5X on Intel’s D-series.
Nvidia dominates data center AI deployments, so challengers must offer buyers a clear reason to change platforms. Intel’s pitch combines high memory capacity with a lower-friction physical format and pairing with its Xeon server processors. The Open Compute Project, where the D-series was shown, is an industry community focused on open data center hardware designs. Source: ‘Put every prompt in its place’: Intel launches 480GB GPU D-series for AI inference – TechRadar, a report on the memory, performance and deployment specifications of Intel’s new AI inference accelerator.Sources