Intel Details Diamond Rapids Xeon and New AI Accelerators for HPC
Intel outlines forthcoming Xeon server chips, a GPU accelerator line and optical links aimed at scientific and AI workloads, according to TechRadar.

Intel has set out a roadmap for high-performance computing that pairs upcoming Xeon server chips with dedicated AI accelerators, optical connectivity work and quantum research, according to TechRadar. The company says the plan is meant to serve ever more demanding scientific workloads while letting research institutions keep the software and systems they already run.
At the center of the hardware push is Diamond Rapids, the next generation of Intel server processors. The design raises capacity through more cores, a larger cache and much greater memory throughput.
Intel says the architecture can be configured with as many as 256 performance cores, memory bandwidth of 1.6 TB/s and 1.28 GB of last-level cache. It also brings Advanced Performance Extensions, AVX 10.2 and a refined version of Advanced Matrix Extensions, which together add hardware acceleration for heavy scientific calculations and AI work.
Diamond Rapids is positioned as an extension of the Xeon 6 line, which Intel says handles intensive simulation, mathematical modeling and other compute-heavy tasks. Jeff McVeigh, who leads Data Center Strategic Programs at Intel as a vice president and general manager, describes the company's stance as favoring cooperation among hardware and software layers over reliance on one architecture.
Alongside the CPUs, Intel is developing a GPU D-series for research groups that need large accelerator memory plus AI processing and double-precision math. The cards will carry up to 480GB of LPDDR5X memory and support numeric formats from FP4 and MXFP4 through to FP64.
Intel says the accelerators use standard PCIe connections and air cooling, so they can be installed in existing facilities without special cooling systems. More memory on board could cut the back-and-forth transfers between processors and accelerators when scientific code works with large datasets or demanding AI models.
Software remains part of the effort. Intel's toolset covers MPI, Fortran compilers, optimization utilities and oneAPI, which lets parallel code run across varied hardware setups. McVeigh argues that faster chips matter little if developers must abandon working applications or rewrite large amounts of code, so Intel plans to have software ready as hardware ships, including support for common AI frameworks and inference environments.
Intel is also working on silicon photonics, which carries data as light rather than electrical signals. The company says this could lower energy use and delays as links stretch between processors, memory and accelerators in large supercomputing sites. Quantum computing is another area of research.
Based on reporting by TechRadar. RoxMag adds plain-English context and checks every figure against that source.
intelxeonhpcai acceleratorssilicon photonics
More in Computing

IBM's Nighthawk r3 to pack 520 qubits on road to 2027 system
IBM has detailed a 520-qubit processor due in 2027, when it plans to pair two chips into a system exceeding 1,000 qubits.

Envision says 800V DC power system running at Inner Mongolia AI site
Chinese firm Envision claims the first commercial deployment of an 800-volt DC power architecture at its Galaxy AI campus in Ulanqab, Inner Mongolia.

ASUS launches full-stack agentic AI offering for enterprise datacenters
ASUS unveiled a three-part Enterprise Agentic AI Solution at NVIDIA's GTC event in Berlin, aimed at companies building AI infrastructure in-house.

Chinese Researchers Report 3.2 Pb/s Data Transfer Over 24-Core Fiber
A Chinese research collaboration says it moved data at 3.2 petabits per second through a 24-core fiber, beating its own earlier record.