Intel has given its Crescent Island inference accelerator a product name, the GPU D-series, and attached figures to it: up to 480GB of LPDDR5X, 1.5TB/s of memory bandwidth and 1.3 petaflops of FP8 compute inside a 350-watt envelope, according to Intel’s announcement at the OCP Global Summit. The card is built on the Xe3P architecture in a standard air-cooled PCIe form factor, and customer sampling is under way. Intel gave no price, launch date or inference benchmark.
The 480GB belongs to partner cards, while the card Intel brands itself has 160GB
A footnote in the release states that the PCIe card Intel brands itself carries 160GB of LPDDR5X, and that the design lets partners build ODM cards under their own brand with options up to 480GB. The capacity in the headline is therefore the top of the range partners can build, not what Intel’s own card holds.
Intel sizes the platform at models of up to 300 billion parameters on a single card and more than one trillion parameters in an eight-GPU system, and it does not say which precision it assumes. At FP8, one byte per parameter, 300 billion parameters is 300GB of weights, which is nearly twice the 160GB on Intel’s own card and likely fits only on the larger partner configurations. Lower-precision weights would shrink that, but Intel does not specify them.
At 1.5TB/s a model that fills the card is read through about three times a second
Reading all 480GB once at 1.5TB/s takes about 0.32 seconds, so a dense model that fills the card may generate no more than roughly three tokens per second for a single request, and a 300GB FP8 model may top out near five, because every token has to stream the weights from memory. That is our arithmetic from Intel’s figures, not an Intel number, and it describes one stream: batching many requests shares each pass over the weights, and a mixture-of-experts model reads only part of them per token.
Intel pitches the card for large language models, long-context work and concurrent inference services. Its figures show why batching matters: 1.3 petaflops against 1.5TB/s is about 870 operations per byte read, so the compute likely stays busy only when each weight fetched is reused across a large batch.
The software is the existing Xe stack, with day-zero support for Meta’s Muse Glimmer
Intel describes the D-series as running on the common Xe software foundation shared with its other GPUs, with upstream framework integration. It lists day-zero support for Meta’s Muse Glimmer on both the D-series and Xeon processors. The pitch is an air-cooled card that fits existing racks without specialized cooling, with OEMs, ODMs and neoclouds as partners, though Intel does not identify any of them.