Intel Unveils AI Inference GPU Loaded With Low-Power DRAM Instead of HBM

Intel's D-Series GPUs offer a distinct approach to AI inference by prioritizing memory capacity and cost-effectiveness over peak memory bandwidth. This choice, which involves using LPDDR5X DRAM, appears to address a market segment where HBM supply has tightened and costs have risen. The product is designed for corporate buyers needing extensive memory for tasks like long-context processing and in-house enterprise AI, particularly in environments with existing air-cooled servers.

The D-Series is built on Intel's Xe3P architecture and can carry up to 480 gigabytes of LPDDR5X DRAM, a capacity increase intended to better suit the analysis of long documents or code. While memory bandwidth is 1.5 terabytes per second and computing performance reaches 1.3 petaflops in FP8 operations at 350 watts, analysts note that the lower bandwidth of LPDDR5X imposes limits compared to HBM. Intel's strategy seems to be absorbing some demand for inference GPUs that do not require the most expensive memory or highest-performing solutions.

This product launch indicates Intel's continued focus on AI accelerators, following previous announcements such as the Crescent Island server GPU. Anil Nanduri, who oversees AI products for Intel's data center group, stated that the GPU D-Series enables practical and cost-effective large-scale inference. This move suggests Intel is positioning itself to meet varied enterprise AI needs, acknowledging that not every workload requires the highest-tier memory performance.