Intel has unveiled the GPU D-series at the Open Compute Project Global Summit, an accelerator with up to 480GB of LPDDR5X memory, more than many laptop SSDs hold. Formerly known as Crescent Island, the chip is designed for AI inference, the stage where already trained models run, rather than for gaming.
The most surprising choice concerns the memory type. Instead of the GDDR7 used in modern graphics cards, Intel chose LPDDR5X, the technology typical of portable and low-power devices. This favors capacity and energy efficiency over raw speed, and points to the goal of hosting very large models on a single accelerator.
According to Intel, one card can handle models with up to 300 billion parameters, while a system with eight GPUs exceeds one trillion parameters. The comparison with Nvidia is immediate: the B300 Blackwell Ultra offers 288GB of HBM3e memory, so the D-series surpasses it in per-GPU capacity, per figures reported by TechRadar.
Capacity alone does not tell how fast the chip is. Intel rates it at 1.3 petaflops in FP8, a numerical format used to speed up AI while reducing memory needs, and 1.5 TB/s of memory bandwidth. These figures measure different things: how much data the chip can hold and how quickly it moves that data to the processor.
The accelerator is designed to fit existing data centers. It comes as a standard PCIe card, drawing 350 W with air cooling, without requiring special cooling systems or new racks. Intel also plans to pair it with its Xeon processors, which handle the stages of the workload that do not need the GPU.
Intel describes an architecture for multi-stage AI, covering retrieval, reasoning, external tools and execution. Vice president Anil Nanduri said that "inference is becoming the defining AI workload". According to the sources, the card is being sampled to customers, but neither a retail price nor a general availability date has been disclosed.
No comments yet. Be the first!