Details have emerged about the data center-focused Intel Xe-HP Arctic Sound GPUs.
Data center GPU models based on the Xe-HP architecture are expected to be released by Intel in the coming months. Recently, Igor’s Lab released images of Intel’s Xe-HP GPU-based compute cards, codenamed Arctic Sound, and revealed some preliminary features.
The single-piece Intel Arctic Sound 1T features an Xe-HP GPU with 384 EU as well as 16GB of HBM2E memory offering the highest bandwidth of up to 716GB / s. The Accelerator is a short single slot full height board rated for 150W TDP.
The Arctic Sound 2T card carries a two-tiled Xe-HP GPU, 960 EU (480 × 2) and 32GB of HBM2E DRAM. The accelerator uses a full length full height (FLFH) form factor and is rated for 300W TDP.
The Intel Xe-HP architecture is very different from the company’s Xe-LP architecture as we know it from Iris Xe consumer grade GPUs. The Xe-HP card supports more floating point formats, more process specific instructions, DP4A convolution instruction for deep learning, and Intel’s XMX extensions.
Data center oriented Xe-HP GPUs use completely new execution units with various IPC enhancements. GPUs with HBM2E memory support are made using Intel’s performance optimized 10nm SuperFin process technology.
Intel now allows some of its customers to preview Arctic Sound computing cards that carry single and double-deck Xe-HP applications. Intel’s Xe-HP plans are not entirely clear at this point as the company never detailed them. Meanwhile, the EU count of these two cards is slightly lower than expected. There is no information about the configurations Intel plans to ship these cards in.