Intel to sell Crescent Island AI GPU with up to 480GB DRAM

By Marcus Bennett

Intel will sell the Crescent Island AI inference GPU as a standalone chip, allowing custom server graphics cards to carry as much as 480GB of DRAM.

Intel’s own card design has 160GB of DRAM. The 480GB custom configuration triples that capacity. Server makers can size memory around each AI service. They can build the boards instead of accepting one fixed configuration.

Crescent Island is Intel’s next-generation graphics processing unit for AI inference. Intel will sell the chip independently of a finished graphics card. Customers can attach it to boards designed for particular workloads.

What does 480GB of DRAM change for AI inference?

It changes how much processed information an AI system can keep in DRAM. When a system analyzes long documents or sustains a conversation, it stores previously processed information in DRAM and reuses it. The stored data forms a key-value, or KV, cache.

If DRAM capacity runs short, part of that cache may be discarded. The system must recompute missing information when needed again. That work can slow response times. Memory requirements therefore vary by AI service, even when the same GPU handles each workload.

Intel adopted LPDDR5X to provide large memory capacity while controlling power consumption. This low-power DRAM appears in smartphones and other devices. Intel selected it for Crescent Island.

The standalone sales model gives server makers control over this configuration. Customers can match DRAM capacity to the memory demands of the service. That choice can reduce cache losses and avoid repeated computation. It also gives each board designer responsibility for memory sizing.

Samples, suppliers and market impact

Anil Nanduri heads AI product and go-to-market strategy in Intel’s Data Center AI Group. He outlined the strategy on the 12th at Intel AI Summit Seoul 2026. The event took place at COEX in Seoul’s Gangnam district.

Nanduri said: “Within weeks of getting physical samples of Crescent Island, we succeeded in running AI inference workloads in our own labs. We are currently shipping evaluation samples to customers.”

The statement places Crescent Island at the customer-evaluation stage. Intel has tested the chip in its own laboratories and has begun shipping evaluation samples. The standalone format makes each customer responsible for board design and memory sizing.

Samsung Electronics (005930.KS) and SK hynix (000660.KS) already mass-produce low-power DRAM for AI servers. Intel’s push into standalone inference GPUs is expected to widen supply opportunities for both companies.

Nanduri framed the commercial test around AI agents. He said: “When companies adopt AI agents, they have to weigh cost, performance and response time together. Crescent Island can be competitive in areas where the balance among those three factors matters.”

That balance explains the emphasis on memory. More DRAM can reduce cache losses and repeated computation. LPDDR5X can lower power demand while retaining high capacity. Both factors affect cost and response time. Flexible configurations let server makers balance memory needs against performance.

For buyers, the practical change is control over memory sizing. The decision depends on which AI services will run and how much KV cache data each must retain. The standalone chip shifts that choice from Intel’s fixed card design to the server maker’s board design. The 480GB ceiling sets the upper limit; workload requirements determine the configuration.