SILICON VALLEY — Intel has unveiled its D-Series graphics processing units for artificial intelligence inference, loaded with large amounts of low-power DRAM. Attention is turning to whether low-power DRAM can take the place of high-bandwidth memory, though analysts say the lower bandwidth imposes limits.
Intel disclosed detailed specifications for the D-Series GPUs at the Open Compute Project Global Summit 2026, held on the 12th in San Jose, California, the company said on the 19th.
Intel focused on the point that not every inference workload requires the most expensive memory and the highest-performing GPU. The product appears aimed at corporate buyers who prioritize capacity and ease of installation over peak performance, for uses such as long-context processing, in-house enterprise AI and existing air-cooled servers. It cannot fully replace HBM, but it can absorb some of the demand, analysts said.
Accordingly, the D-Series departs from typical GPUs by making HBM optional and adopting low-power DRAM instead. Intel is seen as having chosen LPDDR5X as the HBM supply crunch worsened and costs soared. In a May report, the technology outlet Tom's Hardware said low-power DRAM would significantly reduce memory bandwidth but that Intel appeared to have made the choice to cut costs.
The D-Series is built on Intel's latest Xe3P architecture, the underlying chip design, and can carry up to 480 gigabytes of LPDDR5X low-power DRAM. The capacity was increased to make the chips better suited to analyzing long documents or code. Memory bandwidth is 1.5 terabytes per second, and computing performance reaches 1.3 petaflops in FP8, or 8-bit floating point, operations at 350 watts. One petaflop refers to 1 quadrillion floating-point operations per second.
"With the GPU D-Series, large-scale inference can be built practically and scaled cost-effectively," said Anil Nanduri, who oversees AI products and go-to-market for Intel's data center group.
Intel had appeared to be abandoning its GPU business after discontinuing the Arc A750 Limited Edition, its flagship graphics card. But it has signaled continued ambition in AI accelerators, announcing Crescent Island, a data center server GPU specialized for inference, at the OCP summit in October last year.