Intel has unveiled its new D-series GPU, formerly known as Crescent Island, designed specifically for server-side AI inference tasks. The hardware was showcased at the Open Compute Project global summit. The company claims that a single accelerator can handle models with up to 300 billion parameters, while an eight-unit cluster supports models exceeding one trillion parameters.
The accelerator utilizes LPDDR5X memory, a departure from standard HBM or GDDR7, to optimize power efficiency. This configuration allows large neural network weights to remain in accessible memory during inference. The device features a performance rating of 1.3 petaflops in FP8 calculations and a memory bandwidth of 1.5 TB/s, with a power consumption limit of 350 watts.
The D-series GPU is built in a standard PCIe form factor and supports air cooling, which is intended to facilitate integration into existing server infrastructure. Intel plans for the units to operate alongside Xeon processors for task distribution and data management. The company has not yet disclosed pricing or a specific release date for the new hardware.
