Skip to the stories
Windows & PowerShell

Intel Diamond Rapids Xeon Claims 256 Cores and 1.6TB/s, With Key Specs Still Unknown

Intel's SC26 preview of Diamond Rapids promises 256 P-cores, 1.6TB/s of memory bandwidth and 1.28GB of last-level cache, but no launch date or memory details yet.

Intel Diamond Rapids Xeon Claims 256 Cores and 1.6TB/s, With Key Specs Still Unknown

Intel used SC26 to put Diamond Rapids, its next-generation Xeon server architecture, at the centre of its high-performance computing pitch. The headline numbers are large: up to 256 performance cores, up to 1.6TB/s of memory bandwidth and up to 1.28GB of last-level cache. Jeff McVeigh, vice president and general manager of Data Center Strategic Programs, laid out the roadmap, which also covers the Intel GPU D-series, HPC software, silicon photonics and quantum work. Intel gave no launch date. That's the first thing a buyer will notice, and it means these are planning numbers, not purchasing numbers.

Intel Diamond Rapids coverage art: an Intel illustration of server processor packages on a stylized circuit board

What Intel is claiming for Diamond Rapids

Intel describes Diamond Rapids, which it detailed at Hot Chips, as an 18A-P design. The pitch is higher compute density, more memory bandwidth, larger caches and expanded CPU acceleration. Intel Advanced Performance Extensions (APX) target per-thread performance. The 1.6TB/s figure comes from what Intel calls a unified memory fabric.

Here's the thing: the number that matters most to a systems architect is the one Intel didn't give. It hasn't said how many memory channels the chip has, what memory speed sits behind the bandwidth claim, or whether 1.6TB/s applies per socket. Without those, you can't turn the figure into a sizing estimate for a memory-bound workload. Treat it as a headline until the datasheet arrives.

The instruction set story is more concrete. The architecture supports Intel AVX 10.2 and enhanced Intel Advanced Matrix Extensions (Intel AMX). Intel calls the new Xeon the only x86 server CPU with built-in matrix compute for HPC and AI workloads. That's a marketing framing, and it's worth checking against what your application stack actually calls. Software that already targets AMX will benefit most, while code built around other accelerators won't see the change.

On raw core count, the comparison Intel invites is with AMD. The top AMD EPYC Venice parts we reviewed carry 256 Zen 6c cores per socket. So 256 P-cores puts Diamond Rapids at parity on headline count, and the competitive argument will rest on per-core performance, memory behaviour and platform cost rather than core totals.

A heterogeneous model for HPC

McVeigh framed the roadmap around mixed systems, where CPUs orchestrate workflows and run parallel communication libraries while accelerators take on the AI work. He wrote that the future of HPC is “open, heterogeneous and balanced.” That's a fair description of where most new supercomputing procurements are heading, and it explains why Intel is pairing a CPU roadmap with a GPU, software and networking story in the same keynote.

For a site planning a refresh, the practical question is how the CPU, the accelerator and the interconnect will be sold together, and whether the software stack lets you mix vendors without rewriting code. Intel's answer leans on oneAPI and standard toolchains, which brings us to the GPU.

Intel GPU D-series and the SC26 demos

The Intel GPU D-series pairs AI acceleration with double-precision (FP64) compute on an air-cooled PCIe card. Datatype support runs from FP4 and MXFP4 through FP64. The Intel-branded GPU D270 carries 160GB of LPDDR5X and 1.5TB/s of memory bandwidth, and partner-designed configurations scale to 480GB.

Intel GPU D270 air-cooled PCIe accelerator render with Intel branding, the D-series card shown in demos at SC26

Intel's argument for research sites centres on memory capacity. Keeping larger working datasets on the card cuts down on partitioning and host-to-accelerator transfers, which are often what slows scientific workflows. Intel also pitches the card for always-on AI agents that generate code and launch compute jobs. That use case is less familiar to most HPC shops, and it's one to watch for how it shapes scheduling and security policy.

Two live demos ran at Intel's SC26 booth. The first is a seismic imaging mini-application from Veridien, ported from HIP and CUDA to SYCL in less than a month. Intel says 90% of its kernels ran well without Intel-specific optimisation. That's a porting anecdote, and the number describes one application, but it's a useful data point for teams weighing how much CUDA-specific code they carry. The second demo splits an inference pipeline between D270 cards handling prefill and SambaNova SN50 units handling decode, with Xeon 6 and Xeon 6+ processors orchestrating.

The benchmark numbers, and what they don't tell you

Intel compared the D270 against NVIDIA's RTX PRO 6000 Blackwell Server Edition. On FP64 financial workloads it claims up to 5.9x the throughput on European Monte Carlo, 5.6x on European Black-Scholes, 4.6x on binomial pricing and 1.92x on American Black-Scholes. The D270 is rated at 350W and the NVIDIA card at 600W. For inference, Intel claims up to 29% lower latency on Drex v2.0 and 1.12x the output token throughput on Qwen3-30B-A3B at FP8.

These are vendor results on pre-production silicon, and we didn't run or audit them. The financial test configuration lists eight NVIDIA cards against a single D270, and Intel doesn't say whether the results are normalised per card. The NVIDIA inference runs used an AWS virtual machine, which adds a variable of its own. If you're building a shortlist, ask for the full test configuration, the software versions and per-card figures before you put any of these ratios in a business case.

Software and interconnect

On software, Intel says it will keep investing in Intel MPI, Fortran, compiler optimisations and oneAPI, so that existing parallel codebases keep performing on mixed CPU and GPU systems. It's also working on upstream integration with AI frameworks, inference runtimes and deployment environments. The stated goal is for standard models and tools to run on new hardware such as the D-series from the first day it's deployed. For most HPC teams, that's the claim that decides whether the hardware is usable at all, so it deserves a test on your own code before any purchase.

Intel also pointed to its silicon photonics work, which uses light to move data over longer distances at what the company describes as a fraction of the energy of copper links. Intel attached no figures or product timeline to that claim, so for now it's a direction rather than a spec.

Quantum and the Classiq deal

On quantum, Intel restated its case for the planned purchase of Classiq. The deal would combine Classiq's automated quantum algorithm synthesis platform with Intel's silicon spin-qubit technology and chip manufacturing. Intel has said it expects the deal to close in the first half of 2027, pending approval. Quantum is still a long-range item for most data centre planners, but the software layer is the part most likely to touch existing HPC workflows first.

What to watch next

My read is that the Diamond Rapids numbers will hold up or fall apart on three details Intel hasn't released yet: memory channel count and speed, per-socket versus aggregate bandwidth, and the launch timeline. Until those arrive, the sensible move for admins is to track the Hot Chips material and the Intel keynote handout, and to test your critical HPC and AI code against oneAPI and the new AMX and AVX 10.2 paths on hardware you already have. The D270 benchmarks are worth revisiting once independent testers can run them on production configurations.

Source: Harold Fritts · www.storagereview.com

Same flag