At the Open Compute Project (OCP) conference in 2026, Intel announced that its Crescent Island AI accelerator will be sold as the Intel GPU D-Series. The company also released its first official performance numbers: a theoretical peak of 1.3 petaflops (PFLOPS) using FP8 precision and a memory bandwidth of 1.5 terabytes per second (TB/s). The D‑Series will ship as an air‑cooled PCIe card with a 350‑watt power budget and up to 480 GB of LPDDR5X memory.
What the specs mean
The 1.3 PFLOPS FP8 figure is a measure of raw compute speed for low‑precision tasks, which are common in AI inference. FP8 – an 8‑bit floating‑point format – lets the chip process more operations per second while using less power than higher‑precision formats. The 1.5 TB/s bandwidth describes how fast data can move between the chip’s memory and its compute units, a key factor for handling large AI models.
Why it matters
For companies building AI services, the new D‑Series promises faster inference at lower power, which could lower operating costs. Developers will need to adapt their models to FP8 to fully benefit, and the high power draw means the cards require robust cooling and power supplies.
