Intel released new details on GPU D270 on October 12, a data-center accelerator designed to run AI models. Previously called Crescent Island, the product emphasizes memory capacity and deployment in existing servers. The company says customer sampling has started. This is not yet a general-availability announcement, and it includes neither a price nor a date when any customer will be able to order a system.
The headline specification is up to 480 GB of LPDDR5X memory, but a footnote changes the picture: Intel’s branded PCIe card has 160 GB. The design lets partners build cards with up to 480 GB. Anyone comparing purchasing proposals therefore needs to check the actual model and configuration, rather than assume every D270 card carries the maximum capacity.
More room for models and long conversations
Intel says the D-series uses Xe3P architecture, with 1.5 TB/s memory bandwidth and 1.3 petaflops of peak FP8 compute within a 350-watt power envelope. These are manufacturer specifications. By themselves, they do not measure how many useful answers a particular service will deliver, how long users will wait, or what each completed task will cost.
Memory matters because model weights are not the only thing occupying it. Long conversation contexts and intermediate information retained during inference also need capacity. Requirements accumulate when several agents or users run concurrently. More memory may reduce the need to split work across cards, but the result depends on model size, numerical precision and how the service manages requests.
What to check before changing infrastructure
Intel presents D270 as an air-cooled PCIe 5.0 card intended for existing data-center environments. The idea is to add inference capacity without designing around a dedicated liquid-cooling system from the outset. Server compatibility, power supply and airflow still need checking. A card that physically fits is not necessarily a system ready for every workload and software combination.
For development and infrastructure teams, the meaningful test is their own model with context lengths and user counts representative of the service. Measure time to first token, answer-generation speed, memory utilization and whole-system power consumption. This also reveals whether moving from another software environment requires changes that erase some expected savings or add maintenance work for the team.
The update points in a clear direction: not every AI service needs the infrastructure used to train its model. For organizations running models continuously, capacity, electricity and software may matter as much as peak compute. D270 is entering customer evaluation; purchasing decisions will need available configurations, prices and real-task performance, beyond the largest number on the memory specification.