MYBIGGAMING
LIVE
Hardware 17 November 2026 • 10 min read •

Intel Outlines Next HPC Step: Xeon, GPU D-Series, and Open Software

Intel presents a heterogeneous vision for HPC and AI, centering on Xeon 6, the new GPU D-Series, and a shift toward open software ecosystems.
Author: Гика • PC •
Intel Outlines Next HPC Step: Xeon, GPU D-Series, and Open Software

High-Performance Computing (HPC) and Artificial Intelligence (AI) are increasingly converging in scientific and industrial applications. Simulations generate data, AI models assist in its analysis, and automated tools can derive new computational steps from it. Intel is addressing this convergence with a system-wide HPC approach: CPUs, GPUs, software, and high-speed interconnects are designed to handle different tasks, ensuring research teams are not locked into a single architecture or a proprietary programming model. The core of this strategy involves Xeon processors and the GPU D-Series, complemented by open software, optical data links, and long-term advancements in quantum computing.

At the Supercomputing 2026 conference in Chicago, Intel presented a concept where the key factor is not just peak FLOPS or memory bandwidth, but the efficiency of data movement. Modern tasks—ranging from physical simulations to AI-driven data analysis—require a distributed allocation of resources. In this model, the CPU acts as an orchestrator, managing workflows and communication libraries like MPI (Message Passing Interface), while accelerators handle the most computationally intensive mathematical operations. This heterogeneous approach aims to balance raw computing power with memory access, data transfer, and energy efficiency under real-world operating conditions.

Xeon: The Foundation for Heterogeneous Workloads

The Xeon 6 family serves as the CPU base for parallel scientific applications and the orchestration of heterogeneous workflows. Intel is particularly highlighting the upcoming Diamond Rapids architecture, which, according to manufacturer claims, is designed to increase compute density while providing higher memory bandwidth, larger caches, and additional CPU acceleration. While specific release dates and independent comparative measurements have not been provided, Intel positions Diamond Rapids as a significant leap for data center workloads.

Key technical specifications attributed to the Diamond Rapids architecture include up to 256 performance cores and improved single-thread performance supported by Intel Advanced Performance Extensions (APX). Intel also claims a unified memory subsystem with up to 1.6 TB/s of bandwidth and a Last-Level Cache (LLC) of up to 1.28 GB. To accelerate vector and matrix operations essential for scientific computing and AI, Intel is integrating AVX 10.2 and expanded Intel Advanced Matrix Extensions (AMX). The company positions this upcoming generation as the only x86 server CPU with integrated matrix computation specifically for HPC and AI workloads.

GPU D-Series: Balancing AI Inference and FP64 Precision

The Intel GPU D-Series is primarily designed for AI inference—the execution of already trained models. The D270 model is intended to cover selected HPC and mixed AI+HPC tasks. Built on the Xe-3P architecture, it features 32 Xe cores and 256 Xe Matrix Extensions (XMX) to accelerate matrix operations. A notable feature is its support for a wide range of data types, from low-precision formats like FP4 and MXFP4 to FP64 (double precision), which remains critical for numerically demanding simulations and modeling.

  • Architecture: Xe-3P with 32 Xe cores and 256 XMX units.
  • Memory: 160 GB LPDDR5X (up to 480 GB in partner configurations) with 1.5 TB/s bandwidth.
  • Precision Support: From FP4/MXFP4 up to FP64.
  • Form Factor: Air-cooled PCIe card with a 350W TDP.

This flexibility in precision allows the D270 to serve diverse roles, from high-precision financial modeling to energy-efficient AI inference. Intel's vision for AI agents in scientific workflows involves these agents using GPUs to handle the compute-intensive 'prefill' stage (processing input) while specialized accelerators, such as SambaNova's RDUs, handle the 'decode' stage (generating responses), with Xeon CPUs managing the overall orchestration. This disaggregated approach aims to optimize resource usage based on the specific phase of the AI workload.

Software Ecosystem and Connectivity

To avoid vendor lock-in, Intel is emphasizing an open software ecosystem. This includes long-term support for open standards and libraries, ensuring that scientific code can be ported across different architectures without being tied to proprietary models. Furthermore, Intel is exploring silicon photonics to address the growing need for high-speed, low-latency data movement in large-scale clusters, aiming to replace traditional electrical interconnects with optical solutions to improve energy efficiency and bandwidth.

Intel is moving toward a vision of High-Performance Computing (HPC) where the focus shifts from raw peak performance to the efficiency of data movement and software portability. The company's strategy involves a heterogeneous approach, integrating Xeon processors, specialized GPU D-Series accelerators, and silicon photonics, all tied together by an open software stack to prevent vendor lock-in.

Specialized Performance: Financial and Seismic Workloads

In the realm of financial applications, Intel reports that its D270 GPU achieves a throughput 1.92 to 5.9 times higher than the NVIDIA RTX PRO 6000 Blackwell Server Edition. These results were obtained using Binomial, Black-Scholes, and Monte Carlo methods for valuing European and US options. It is crucial to note that Intel emphasizes these are application-specific results rather than a comparison of theoretical peak performance. The tests utilized SYCL/MKL, and Intel selected the best available results from both CUDA and SYCL/MKL for the comparison card. Since detailed test configurations and individual values were not provided, these figures cannot be directly applied to other systems.

Seismic analysis benchmarks also show significant differences. In a MiniMod benchmark involving a 3D grid with one billion points, an optimized SYCL implementation on the D270 reached 45.91 billion grid points per second. In comparison, the Intel Arc Pro B70 achieved 22.52 billion points per second, representing a 2.04 to 1 ratio. While these are manufacturer-provided figures without independent verification, they suggest substantial potential for applications requiring high FP64 (double-precision) throughput.

Software Portability and Connectivity

For HPC users, hardware is only as viable as the software that runs on it. Intel's strategy focuses on an open, unified stack designed for both existing infrastructure and new AI workloads. This includes Intel MPI, Fortran, compiler optimizations, and the oneAPI platform. A key goal is 'Day-0' readiness, ensuring software is immediately functional upon the release of new hardware. For the GPU D-Series, this means early integration with common AI frameworks, inference runtimes, and deployment systems for both CPUs and GPUs, aiming to allow hybrid HPC and AI workflows without forcing users to abandon code portability.

To address the challenges of scaling large clusters, Intel is exploring silicon photonics. This technology uses light instead of electrical signals over copper to transmit data over longer distances, aiming to reduce latency and energy consumption. While Intel has not provided specific measurement values or comparison conditions, silicon photonics is a core part of their long-term infrastructure strategy to bring compute accelerators, memory, and other resources closer together.

Quantum Computing as a Long-Term Complement

Looking further ahead, Intel views quantum computing as a complementary technology rather than a replacement for classical supercomputers. Following the announced agreement to acquire the software company Classiq (expected to close after October 26, 2026), Intel aims to combine Classiq’s automated quantum circuit synthesis platform with Intel's silicon spin-qubit technology. The goal is to create a software layer that abstracts complex quantum physics, making quantum algorithms more accessible to classical HPC developers.

A System-Wide Approach with Open Questions

Intel's vision for the future HPC system relies on complementary resources: Xeon processors for coordination and computation, GPUs for AI and scientific acceleration, open software to bridge workloads, and optical interconnects for efficient data movement. While the logic of distributing tasks to the most suitable components is sound, the actual success of this ecosystem will depend on software support, system integration, and independent application testing. Currently, many of these performance claims, including those for the Diamond Rapids architecture, remain manufacturer-driven targets without independent benchmarks or confirmed release dates.

At the Supercomputing 2026 conference in Chicago, Intel presented a strategic vision for the future of High-Performance Computing (HPC) and Artificial Intelligence (AI). The company is moving away from a single-architecture approach, instead proposing a heterogeneous ecosystem where CPUs, GPUs, and specialized accelerators work in tandem, unified by open software and high-speed interconnects. This shift aims to address the growing complexity of hybrid workloads, where data movement efficiency is becoming as critical as raw floating-point performance (FLOPS).

Xeon: The Foundation for Heterogeneous Computing

The cornerstone of this architecture is the Xeon 6 family, designed to act as the orchestrator for complex workloads. A major focus is placed on the upcoming Diamond Rapids architecture. According to Intel, this architecture will provide significantly higher compute density, an increased Last Level Cache (LLC) of up to 1.28 GB, and memory bandwidth reaching 1.6 TB/s. Intel claims that Diamond Rapids will be the only server x86 platform with integrated matrix computations for AI and HPC tasks, thanks to support for AVX 10.2 and AMX (Advanced Matrix Extensions). It is important to note that independent comparative tests and exact release dates for this architecture have not yet been provided.

To enhance per-thread performance, Intel plans to implement APX (Advanced Performance Extensions). Combined with an increased core count—reaching up to 256 Performance-cores in top-tier configurations—this is intended to provide the necessary scalability for large-scale scientific simulations.

GPU D-Series: Balancing AI Inference and FP64

A new entry in the accelerator segment is the Intel GPU D-Series. The D270 model is primarily positioned for AI inference (running pre-trained models), but it features a critical capability: support for double-precision floating-point (FP64) computations. This makes it suitable for mixed workloads, such as using AI to interpret results from complex physical simulations.

  • Architecture: Xe-3P with 32 Xe cores and 256 XMX matrix extensions.
  • Memory: 160 GB LPDDR5X (expandable to 480 GB in partner configurations) with 1.5 TB/s bandwidth.
  • Data Precision: Support for formats ranging from FP4 and MXFP4 to FP64.
  • Form Factor: Air-cooled, PCIe interface, with a power consumption of up to 350 W.

This flexibility in precision allows the D270 to serve various scenarios, from high-precision financial calculations to energy-efficient AI inference. Intel also demonstrated a concept of task decomposition: for instance, when working with Large Language Models (LLMs), GPUs can handle the compute-intensive Prefill stage, while specialized accelerators (like SambaNova's RDU) or Xeon processors manage the Decode stage.

Performance in Specialized Computations

In the financial sector, Intel claims significant advantages with its architecture. The D270's throughput in option pricing tasks (using Binomial, Black-Scholes, and Monte-Carlo methods) was reported to be 1.92 to 5.9 times higher than the NVIDIA RTX PRO 6000 Blackwell Server Edition. It is crucial to note that Intel is comparing specific algorithmic performance using SYCL/MKL rather than theoretical peak power. While Intel optimized the comparison by using the best available results for both sides (CUDA vs. SYCL/MKL), the lack of detailed testing methodologies makes it difficult to generalize these figures to other systems.

In seismic analysis, Intel presented impressive internal benchmarks. In the MiniMod benchmark on a 3D grid with one billion points, an optimized SYCL implementation achieved 45.91 billion points per second on the D270. In comparison, the Intel Arc Pro B70 delivers 22.52 billion points per second, representing a 2.04x advantage. These are internal company figures that have not undergone independent verification, but they highlight the architecture's potential for FP64-intensive tasks.

Software Ecosystem and Networking

For the HPC sector, compatibility with existing code—often built on decades of specific libraries—is vital. Intel's strategy relies on an open, unified stack including Intel MPI, Fortran, compiler optimizations, and the oneAPI platform. The goal is 'Day-0' readiness: software should be ready to run on new hardware the moment it is released. For the D-Series, this means early integration with popular AI frameworks, allowing users to work in hybrid CPU+GPU environments without being locked into a single vendor's proprietary roadmap.

As cluster scales increase, data transfer becomes a bottleneck. Intel is exploring silicon photonics—using light instead of electrical signals over copper cables—to reduce latency and power consumption in large-scale systems. Specific technical parameters and testing conditions for this technology remain undisclosed.

Quantum Computing and Systemic Integration

In the long term, Intel plans to integrate quantum computing as an adjunct to classical supercomputers. Following the acquisition of Classiq (expected to close after October 26, 2026), Intel intends to combine automated quantum circuit synthesis with its silicon-based spin qubit technology. The idea is to create a software layer that abstracts quantum physics complexity, allowing classical developers to utilize quantum power without deep hardware modeling.

The overall vision for Intel's HPC future is a symbiosis: Xeon processors coordinating processes, GPUs accelerating AI and scientific tasks, open software linking workloads, and photonics optimizing data movement. However, the actual realization will depend on developer adoption and independent validation, which is currently lacking for many key components like the Diamond Rapids architecture.

Tags
Share
Article author

Гика

Cookies

We use cookies

The site uses technically necessary cookies for sign-in, security, storing preferences, and correct page operation. Analytics and other non-essential technologies should only be enabled after your consent. You can read more about data categories, purposes, and retention in the privacy policy.

Cookie policy
Quick actions