Intel reportedly shared further HPC details at SC26 during the week of November 17, 2026, according to MuyComputerPro. Separately, Intel’s Hot Chips 2026 handout describes Diamond Rapids as a next-generation Xeon CPU, and its product page specifies the D270 as an inference-optimized data-center GPU.

In HPC, a CPU’s core count is just one part of the picture: memory capacity and bandwidth, cache, and configurable I/O also shape how it fits into a larger system.

What Intel specifies for Diamond Rapids

Intel’s Hot Chips handout lays out a modular design with four compute building blocks, 16 core chiplets, four base tiles, and two Fabric Hub tiles. Each core chiplet supports up to 16 cores, giving Intel’s stated maximum of 256 cores in total. The handout also lists up to 1.28 GB of last-level cache (LLC) across all compute building blocks. Intel’s Diamond Rapids handout

The package uses Foveros 3D Direct hybrid bonding. Intel identifies substrate copper links as UCIe-S at 16G and identifies 18A-P for Diamond Rapids. These details describe the package and its connections; the core count alone doesn’t describe the whole platform.

Memory and I/O are configurable parts of the design

Intel lists 16 memory channels, supporting DDR5 speeds up to 8000 MT/s or MRDIMM speeds up to 12800 MT/s, with memory bandwidth up to 1.6 TB/s. The transfer-rate figures and the bandwidth ceiling describe different parts of the memory specification.

For I/O, Intel specifies 128 flexible lanes that can be configured for PCIe Gen 6, CXL 3.0, UPI 3, or combinations of those interfaces. The processor also has eight PCIe Gen 4 lanes, and Intel lists up to 2 TB/s of bidirectional I/O-fabric bandwidth. The 128-lane figure is a configurable pool, not 128 lanes for each interface.

The handout also lists APX, AVX 10.2, AMX, TDX, TDX Connect, SGX, QAT, DSA, and IAA. That mix of instruction-set extensions, security features, and accelerators complements the processor’s memory and I/O resources in Intel’s stated platform design.

The D270 has a different HPC role

Intel describes the D270 as an inference-optimized data-center GPU. Its product page specifies up to 480 GB of LPDDR5X per card, 350 W of air-cooled operation, PCIe 5.0, 32 Xe3P cores, and 256 XMX matrix engines. Intel positions the card for production inference, rather than training models at massive scale. Intel’s D-series product page

That distinction matters when mapping the parts to a workload: Diamond Rapids is the CPU architecture, while Intel presents the D270 for inference tasks. The D270’s listed specifications belong to that card, not to the D-series family as a whole.

What Intel’s FP64 comparison says

MuyComputerPro attributed to Intel a claim of up to almost six times the performance of the NVIDIA RTX PRO 6000 Blackwell Server in some native FP64 workloads. Intel’s product page describes its RTX PRO 6000 projections as based on publicly available information and Intel’s internal analysis as of February 2026. The claim is limited to some native FP64 workloads; it is not a general result for every HPC task. (MuyComputerPro’s report; Intel’s D-series page)

How Intel describes a heterogeneous HPC workflow

Intel illustrates its CPU-and-accelerator approach with a structural-simulation workflow. An AI coding agent generates MPI code; the CPU orchestrates the workflow and communication libraries, accelerators handle compute-intensive AI tasks, and interconnects move data. Intel’s explanation of its HPC approach

The example assigns different jobs to the parts of a system rather than treating the CPU and accelerator as interchangeable. It is Intel’s illustration of a workflow, not a Diamond Rapids performance measurement.

What timing has been reported for Diamond Rapids?

ServeTheHome reported a 2027 roadmap target for Diamond Rapids and described a second-half arrival as likely. Intel’s handout says codenames refer to products in development and not publicly available. ServeTheHome’s report