
Intel Xeon and GPU D-Series: HPC and AI in a Unified System - igor´sLAB
Intel has introduced a heterogeneous computing approach combining Xeon processors and GPU D-Series accelerators with open software and high-speed connections for HPC and AI workloads.
At Supercomputing 2026 in Chicago, Intel is presenting a system-wide approach to high-performance computing (HPC) and artificial intelligence that combines CPUs, GPUs, software, and high-speed connections. The company's Xeon processors and GPU D-Series accelerators are designed to handle different tasks without locking research teams into a single architecture or proprietary programming model. This heterogeneous computing approach recognizes that different phases of a workflow require different types of processors.
Intel's concept moves beyond traditional metrics like peak FLOPS (floating-point operations per second) and memory bandwidth, emphasizing instead how computing performance, memory access, data transfer, and software work together efficiently under real-world constraints on energy, heat generation, and cost. The company cites an example of an AI coding agent that generates MPI code for structural simulations, where CPUs coordinate workflows while accelerators handle compute-intensive tasks.
Xeon as Host Platform
The Intel Xeon 6 processor family provides the foundation for parallel scientific applications and orchestrating heterogeneous workflows. Intel highlights performance cores and support for high-bandwidth memory. The upcoming Diamond Rapids architecture promises increased compute density with up to 256 performance cores, higher performance per thread, and Intel Advanced Performance Extensions (APX).
Diamond Rapids will feature a unified memory fabric with bandwidth up to 1.6 TB/s and up to 1.28 GB of Last-Level Cache (LLC). Intel also lists AVX 10.2 and extended Intel Advanced Matrix Extensions (AMX) as integrated features that accelerate vector and matrix operations used in scientific computing and AI applications. The company positions this Xeon generation as the only x86 server CPU with integrated matrix computing for HPC and AI workloads.
GPU D-Series for AI and HPC
The Intel GPU D-Series is primarily designed for AI inference but can also handle selected HPC and mixed AI+HPC tasks. The D270 model features Xe-3P architecture with 32 Xe cores and 256 Xe Matrix Extensions (XMX), supporting data types from FP4 and MXFP4 to FP64. Intel claims the vector units deliver full-rate FP64 performance with 64 FP64 fused multiply-add operations per Xe core.
The D270 is intended to allow the choice of precision to vary by task, rather than requiring an entire application to use a single number format.
The GPU comes with 160 GB of LPDDR5X memory (expandable to 480 GB in partner configurations) and memory bandwidth of 1.5 TB/s. Designed as an air-cooled PCIe card with 350-watt power consumption, the D-Series is positioned as an inference accelerator rather than a replacement for all traditional supercomputing GPUs.
Performance and Software Ecosystem
Intel cites performance advantages in specific applications, claiming the D270 delivers 1.92 to 5.9 times higher application throughput than an NVIDIA RTX PRO 6000 Blackwell Server Edition in financial services applications using binomial, Black-Scholes, and Monte Carlo methods. In seismic analysis, Intel reports an optimized SYCL implementation achieving 45.91 billion grid points per second in a MiniMod benchmark, compared to 22.52 billion for the Intel Arc Pro B70.
Intel is also developing disaggregated inference architectures in partnership with SambaNova, where GPUs handle compute-intensive prefill, SambaNova RDUs perform decode operations, and Intel Xeon processors orchestrate workflows. The company emphasizes software portability through SYCL and other open standards, allowing applications to run across different hardware configurations without vendor lock-in.