Intel introduced its Crescent Island-based D-series inference GPUs. The D270 specifies up to 480 GB of LPDDR5X memory, a 350 W air-cooled design, and 1.5 TB/s bandwidth, emphasizing support for existing software such as PyTorch. Intel also says a single card can handle a 300-billion-parameter model and an eight-card setup a one-trillion-parameter model. These figures depend on configurations specified by Intel. The products are at the sampling stage, so volume availability and real service performance are not yet established.
Back to Tech
Intel unveils D-series inference GPUs with up to 480 GB of memory
Model-capacity figures reflect vendor-specified conditions; real serving latency and costs require separate verification.