Intel's new D270 AI card packs up to 480 GB of memory

For most of the AI boom, the story of data center chips has been a story about speed. Faster memory, faster links, more petaflops, more power per rack, and more liquid cooling to carry the heat away. On October 12, 2026, at the Open Compute Project (OCP) Global Summit, Intel made a different bet. It gave its long teased inference accelerator, code named Crescent Island, a real product name: the Intel GPU D series, with the GPU D270 as its first model.

The headline number is memory. Partner built D270 cards can carry up to 480 GB of LPDDR5X, the kind of low power memory you usually find in laptops and phones, rather than the expensive stacked HBM that sits next to almost every high end AI chip. Intel pairs that with 1.5 TB/s of memory bandwidth and 1.3 petaflops of FP8 compute, all inside a 350 watt, air cooled PCIe 5.0 card that slides into ordinary servers. Intel says customer sampling is already underway. What it did not say is what the card costs or when it ships in volume.

The official announcement is on the Intel newsroom, with more details on the Intel GPU D270 product page. Below is what the card is, why Intel chose capacity over speed, what fits on it, how it compares, and what is still unknown.

What Intel actually announced

The D series is the first product built on Intel's Xe3P architecture. Each D270 has 32 Xe3P cores and 256 XMX matrix engines, the units that do the heavy matrix math behind neural networks. Reporting from Notebookcheck adds a 32 MB unified L2 cache. Intel says it took die area that would normally go to graphics and ray tracing pipelines and gave it to compute instead, so despite the GPU name this is not a card you could game on. It exists to run AI models.

There are two memory configurations. Intel's own branded card ships with 160 GB, while the design lets server makers build their own cards with flexible options up to 480 GB. That distinction matters, because the 480 GB figure is the one that will travel around social media, but it applies to partner hardware, not to the card with the Intel logo.

The rest of the spec sheet is deliberately conservative. 350 watts is modest for a modern accelerator. Cooling is passive and air based, so the card relies on the server's fans rather than a liquid loop. The interface is standard PCIe 5.0, and Intel talks about systems with up to eight cards, much like the familiar eight GPU servers that dominate AI clusters today.

Why capacity instead of speed

To understand the design, it helps to look at what AI agents actually do. Intel's product page puts it in one line: agents read far more than they write. A coding agent or a research agent keeps a long history in its context, often 100,000 tokens or more, and it calls the model again and again over thousands of steps. Every call starts with prefill, where the model processes the whole prompt, and then moves to decode, where it generates the answer token by token.

All of that context has to live somewhere. During inference, a model keeps a KV cache, a memory of the keys and values for every token it has already seen. The cache grows with the length of the context and with the number of users served at once. For long running agents and many parallel sessions, the cache can grow faster than the need for raw compute. When it no longer fits, a provider has to spread the work across more cards, or throw away context, or both.

HBM solves the speed side of this problem beautifully, but it is expensive, supply constrained and tied to advanced packaging. The memory market has been tight all year, and HBM is the most contested part of it. LPDDR5X is cheaper per gigabyte, easier to buy and much more power efficient, but it is also much slower. Intel's argument is that for a growing share of inference work, how much fits on a card matters more than how fast each byte moves. As StorageReview put it, Intel is positioning capacity per watt and per dollar as the figure that matters for decode heavy serving, ahead of peak bandwidth.

How it compares

The most useful comparison is with Intel's own current inference card. Gaudi 3 carries 128 GB of HBM2e at 3.7 TB/s. A 480 GB D270 has close to four times the memory, but well under half the bandwidth. That is the whole trade in one sentence.

On the Nvidia side, Intel's product page says its performance projections are made against the RTX Pro 6000, Nvidia's Blackwell based professional card with 96 GB of GDDR7 and roughly 1.8 TB/s of bandwidth. Against that card, the D270 offers five times the memory at a similar bandwidth level, which is exactly the comparison Intel wants people to make.

What the D270 does not try to do is compete with Nvidia's top systems. Tom's Hardware points out that each Rubin GPU in Nvidia's Vera Rubin NVL72 rack is rated at about 17.5 petaflops of FP8, more than ten times the D270, but needs liquid cooling and draws power measured in kilowatts. These are different products for different buyers. The D270 is aimed at companies that want to add inference capacity to the data centers they already have, not at labs building training clusters.

What fits on one card

Intel says a single D series card can host models of up to about 300 billion parameters, and an eight card system can scale past one trillion parameters. A quick back of the envelope check shows why that works on a 480 GB card.

At FP8 precision, each parameter takes roughly one byte. A dense 300 billion parameter model therefore needs about 300 GB just for its weights. That leaves around 180 GB for the KV cache, longer contexts, multiple users and runtime overhead. On Intel's own 160 GB card, the same model would not fit, so the headline capacity claim clearly refers to the bigger partner versions.

The same math also shows the limit. During decode, a dense model has to read all of its active weights once for every new token. At 1.5 TB/s, reading 300 GB takes about a fifth of a second, which caps a single stream of a dense 300B model at roughly five tokens per second. That sounds slow, and for one user it would be. Real deployments are different in two ways. First, servers batch many users together, so the weights are read once and used for many requests at the same time. Second, most large modern models use a mixture of experts design, where only a fraction of the parameters is active for each token. Both push real throughput far above the naive number. Still, this is the honest picture: the D270 is built to hold a lot and serve many sessions efficiently, not to be the fastest card for a single chat.

These figures are our own estimate based on Intel's published numbers, not a benchmark. Intel has not released throughput results for specific models, and nobody outside Intel and its sampling customers has tested the card yet.

Software and partners

A new accelerator lives or dies by its software, and that is where Intel has struggled before. This time it leans on a common Xe software foundation shared with its integrated and discrete GPUs. According to StorageReview, code developed on Intel's Arc Pro cards is meant to carry over to the D series without kernel rewrites. Intel says the stack integrates upstream with standard runtimes, names vLLM and the open PyTorch ecosystem, and promises day zero support for new open models. Its first example is Meta's Muse Glimmer, a 30 billion parameter open weight model built for local agentic work.

The ecosystem slide Intel showed reportedly ran to 53 logos. Server names include Dell, HPE, Lenovo, Supermicro, Cisco, ASUS, ASRock Rack, Giga Computing, MiTAC, Pegatron, QCT, Wistron, Inventec, Compal and Foxconn. On the software side, StorageReview lists Red Hat, Canonical, VMware by Broadcom, Nutanix, NetApp, WEKA, vLLM, Hugging Face, Anyscale and OpenHands. A logo on a slide is not a purchase order, but it shows that Intel has done the groundwork to make the card easy to try. Phoronix also notes that the Linux driver support has been landing upstream, which should help on day one.

The bigger picture at OCP

The D270 was not Intel's only news. The company also showed Xeon 6+ rack designs, many built to the Open Rack v3 standard, with the densest configurations reaching more than 46,000 CPU cores per rack at 133.45 kW. Intel's pitch is that agents do a lot of work that is not model inference at all: routing requests, running tools, executing code, searching databases and moving data. In Intel's words, the measure of an AI rack is shifting toward the number of concurrent agents it can sustain at a target latency.

Intel was not alone in saying this. On the same day, AMD introduced its Helion rack blueprint, including a CPU only version with up to 32,768 cores aimed at agent execution, and Microsoft detailed its Maia 300 inference chip, which we covered in our Maia 300 article. Three of the largest chip players used the same conference to argue that the next phase of AI hardware is about inference for agents, not only training giant models.

Intel also highlighted its work in the OCP Open Adapter Card workgroup, formed with AMD, Dell, Nvidia and Qualcomm, which is defining a standard form factor for accelerator cards between a plain PCIe card and a full rack system. Intel's Nic Dubé was scheduled to give a keynote on October 13 under the title "Put Every Prompt in Its Place", which sums up the strategy: match each AI task with the cheapest hardware that can do it well.

What we still do not know

There are big gaps in what Intel has shared so far.

• Price. Intel has not given one. For a card that sells itself on cost per gigabyte and per token, this is the most important missing number.
• Launch date. Sampling is underway with server makers, cloud providers and software vendors, but there is no general availability date.
• Real throughput. There are no published tokens per second figures for specific models, and no independent benchmarks.
• Which capacity ships first. The 480 GB versions depend on partners deciding to build them. Intel's own card has 160 GB.
• Track record. Tom's Hardware reminds readers that Intel's earlier data center GPU, Ponte Vecchio, did not take off, and that its follow ups were shelved. Gaudi has also struggled to win share against Nvidia.

Why it matters

For many companies, the bill for running AI now matters more than the bill for training it, and agents make that bill bigger because they keep long contexts and call models over and over. If memory capacity is the real bottleneck for that kind of work, a cheaper, air cooled card with a lot of memory could let enterprises, government clouds and smaller cloud providers run large open models on hardware they can actually buy and power. It also adds a new option in a market where HBM supply and Nvidia pricing set the rules for almost everyone.

For people who care about open models, this is interesting too. A card that can hold a 300 billion parameter model on its own makes it easier to run strong open weight models privately, without splitting them across many expensive GPUs. That is good for data control, sovereignty and competition.

Dany's take

I like the idea more than I trust the timeline. The diagnosis is right: agents are memory hungry, and HBM is the most expensive and most constrained part of the whole AI supply chain. Putting 480 GB of cheap, cool memory on a normal PCIe card is a smart way around that, and it fits the open model wave perfectly. But Intel has announced promising AI chips before and then struggled to ship them at scale or to get developers to care. Without a price, without independent benchmarks and with the biggest memory option depending on partners, the D270 is still a strong slide, not yet a proven product. If Intel lands a competitive price and vLLM support is solid on day one, this could be the first Intel AI card that people actually want. I will believe it when the first independent token numbers show up.

Bottom line

Intel has renamed Crescent Island as the GPU D series and introduced the D270, an inference only card with up to 480 GB of LPDDR5X on partner versions, 160 GB on Intel's own card, 1.5 TB/s of bandwidth, 1.3 petaflops of FP8 and a 350 watt air cooled PCIe 5.0 design. Intel says one card can run models of about 300 billion parameters and eight cards can go past one trillion. Samples are with customers now, but price, launch date and real world performance are still open.

Primary source: Intel newsroom, Intel Accelerates Open Infrastructure for the Agentic AI Era and the Intel GPU D270 product page. Further reporting: StorageReview, Tom's Hardware and Phoronix.