Intel puts Diamond Rapids Xeon memory bandwidth at up to 1.6 TB/s
Editorial Team
••📖6 min read
Intel disclosed the bandwidth target for its next Xeon server processors at SC26, alongside previously announced core and cache limits. Application performance, power requirements and launch timing remain unresolved.
Diamond Rapids will offer up to 1.6 TB/s of memory-fabric bandwidth, Intel says.
Its maximum 256 performance cores and 1.28 GB cache were disclosed before SC26.
Companion GPU tests illustrate Intel’s combined HPC and AI pitch, but do not establish Diamond Rapids performance.
Intel said on November 16 at Supercomputing 2026 that Diamond Rapids, the upcoming generation of its Xeon server processors, will deliver up to 1.6 TB/s of bandwidth through a unified memory fabric. The disclosure adds a key specification to a processor intended to run scientific applications and coordinate systems that combine CPUs with accelerators.
For research institutions, the significance is the balance between computation and data supply. More cores cannot by themselves show how quickly a simulation will finish. Intel is pairing higher core density with more bandwidth, larger caches and integrated acceleration, while leaving the application results needed to evaluate that combination for later.
Company-reported Diamond Rapids ceilings: up to 1.6 TB/s bandwidth, 256 performance cores and 1.28 GB cache. Only the bandwidth figure is newly disclosed at SC26. Source: Intel.
A new bandwidth figure for an existing roadmap
Intel’s SC26 announcement specifies up to 256 performance cores and up to 1.28 GB of last-level cache, alongside the bandwidth ceiling. The cache is intended to keep important datasets close to the cores rather than repeatedly retrieving them from memory.
Those core and cache limits are continuity rather than a fresh unveiling. In its August 24 Hot Chips announcement, Intel had already listed both, as well as 16 memory channels operating at 12,800 MT/s. It also listed 128 PCIe Gen6 lanes and CXL 3.0 support for platform connectivity. That earlier disclosure tied Diamond Rapids to Intel’s own 18A-P manufacturing technology and Foveros Direct 3D packaging. The product therefore carries a stake for Intel’s manufacturing and packaging technologies as well as its server-processor business.
South Korea is strengthening device checks and considering public Wi-Fi shutdowns before its November 19 college entrance exam, after a tax-accountant exam top scorer admitted using AI to cheat. The response still depends heavily on human screening, with no electronic device detectors planned for this year's test.
Editorial Team
Diamond Rapids follows the Xeon 6 generation, which Intel positions for simulation, modeling and the coordination of heterogeneous computing systems. The new architecture adds Advanced Performance Extensions, or APX, and supports AVX 10.2 vector instructions and enhanced Advanced Matrix Extensions, or AMX. These address different kinds of computation; AMX supplies integrated matrix operations used in AI and some scientific workloads.
Intel describes the next Xeon as the only x86 server CPU with built-in matrix compute for HPC and AI. That is a company differentiation claim, not a general performance ranking. On bandwidth, a technical report on the disclosure places the 1.6 TB/s figure level with AMD’s competing EPYC 9006 server-processor family. Matching a headline bandwidth figure does not establish equal application performance.
Intel’s package illustration from its August 2026 Hot Chips announcement, which outlined Diamond Rapids, Crescent Island and Wildcat Lake. Illustration: Intel. Source: Intel Hot Chips announcement.
The companion GPU shows the deployment pitch
Intel’s broader argument is that scientific computing increasingly mixes simulation, analytics and AI, requiring different processors to work together. Its accompanying GPU D270, previously codenamed Crescent Island, is principally an AI-inference accelerator that also supports double-precision scientific calculations.
The D270 uses the Xe3P architecture with 32 Xe cores and 256 XMX matrix engines. The reported specifications put Intel’s card at 160 GB of LPDDR5X memory and 1.5 TB/s of memory bandwidth, with partners able to offer configurations up to 480 GB. Its air-cooled PCIe design is specified at 350 watts. The Xe-core and matrix-engine counts, maximum partner memory capacity and 350-watt design were already disclosed at Hot Chips. Supported formats range from low-precision FP4 and MXFP4 to FP64, or double precision. Intel specifies 64 FP64 fused multiply-add operations per Xe core, but has not supplied clock rates or a theoretical FP64 throughput figure in TFLOPS.
The distinction between Intel’s card and partner configurations matters: the largest memory capacity is not the specification of every D270. Intel argues that retaining larger models and datasets on an accelerator can reduce transfers and partitioning overhead. Air cooling and a conventional expansion-card format are central to its pitch for adding AI capability to existing research infrastructure. Unlike high-end accelerators built around high-bandwidth memory, or HBM, D270 uses LPDDR5X. That is a different memory and deployment design, rather than evidence of a lower total cost of ownership; the disclosures cited here provide no purchase-price comparison.
Intel says in its D270 technical announcement that the application benchmarks used pre-production silicon. Its financial-computing tests show D270 throughput between 1.92 and 5.9 times that of Nvidia’s RTX PRO 6000 Blackwell Server Edition, depending on the application. The workloads cover Binomial, Black-Scholes and Monte Carlo options-pricing calculations. Intel describes the D270 as operating at 350 watts against 600 watts for the Nvidia GPU; those figures do not measure complete-system energy consumption or cost per completed job. These are company tests of particular workloads, not a universal GPU ranking or a Diamond Rapids benchmark. The reported methodology uses SYCL and Intel’s Math Kernel Library on D270, and the better of CUDA or SYCL/MKL results on the Nvidia card. In a separate optimized MiniMod seismic test, a report of Intel’s results gives D270 throughput of 45.91 billion grid points per second against 22.52 billion for Intel’s Arc Pro B70 graphics card, using a three-dimensional grid of one billion points. The reported 2.04-fold ratio therefore concerns that optimized workload and an Intel comparator, rather than the Nvidia card used for the financial tests.
The available inference evidence is narrower than the broad product pitch. One account of Intel’s tests reports up to 29% lower latency in decision-making models and 1.12 times the throughput for Qwen 3 30B-A3B against the RTX PRO 6000 Blackwell. It does not provide enough configuration and measurement detail to turn those ratios into a general tokens-per-second or cost comparison.
Company-reported D270 throughput relative to Nvidia’s RTX PRO 6000 Blackwell Server Edition in four financial workloads. Tests used pre-production silicon; these are GPU results, not Diamond Rapids benchmarks. Source: Intel D270 technical announcement.
Software and launch evidence still to come
Intel is also proposing a divided inference workflow: D270 handles prefill, which processes the input prompt; SambaNova’s SN50 reconfigurable dataflow units handle subsequent decoding; and Xeon 6 and Xeon 6+ processors coordinate work and execute additional tools. Intel plans an SC26 demonstration of that arrangement. It is a proposed division of work across products, rather than evidence of a measured Diamond Rapids system advantage.
Software compatibility is part of the purchasing question. Intel says it will continue investing in MPI, Fortran, compiler optimization and oneAPI to preserve existing scientific code while adding AI. The reported seismic-imaging port offers an early, limited example: a report of Intel’s example says it took less than a month, with 90% of kernels working well without Intel-specific optimization and numerical performance validated. That experience does not establish equivalent portability for every application.
For Diamond Rapids itself, the decisive missing evidence remains application throughput, power consumption and availability. The launch-status report says Intel has not yet supplied a release date, although ASRock has shown a prototype motherboard for Diamond Rapids, also identified as Xeon 7, supporting up to 4 TB of DDR5. Platform preparation is underway; a shipping schedule and comparable system measurements are still needed before buyers can judge what the new bandwidth buys them.
Russia’s objections to observer participation stalled the opening of a Geneva weapons conference, where governments must decide whether to authorize negotiations on autonomous weapons and whether any resulting rules should be binding.
Apple’s iOS 27.3 beta ties higher allowances for server-based AI to existing iCloud+ plans. Basic Siri commands remain free, but undisclosed daily limits leave the value of paid access uncertain.
The partners plan to combine IBM quantum processors, NVIDIA computing and Fermilab physics expertise to study modular quantum systems. A separate decoder release promises faster error-correction research, with important limits on its performance comparisons.
Federal Reserve Chairman Kevin Warsh outlined AI scenarios spanning shared prosperity, job displacement and an investment downturn. The exercise offers no forecast or rate guidance as colleagues worry that the infrastructure buildout is adding to inflation.
OpenAI is offering approved organizations a free classifier for suspected child sexual abuse images, including AI-generated material. Early results suggest less review work, but existing detection alternatives and the need for human review qualify the launch.
Nvidia plans scientific computing racks for early 2027, extending Vera beyond its initial AI deployments. Its 2.4× performance-per-watt claim comes from one airflow simulation against AMD Turin; performance against newer rivals and delivered costs remain unresolved.
OpenAI’s CEO expressed empathy for Nvidia founder Jensen Huang but withheld judgment on reinstatement. His confidence in continued AI financing comes with a direct commercial stake and unresolved deployment risks.
Google’s 8,192-chip Ironwood submission completed the MLPerf DeepSeek-V3 benchmark in 1.70 minutes, close to NVIDIA’s 1.53 minutes at the same accelerator count. Google’s scaling data also shows the extra chip-time required for its fastest result, leaving cost and energy comparisons unresolved.
TensorWave will supply dedicated AMD infrastructure under a six-year OpenAI agreement, with delivery expected from 2027. Its operating record and financial backing sit alongside an undisclosed contract price and a recently terminated site lease.
California alleges OpenAI endangered the public and misled consumers by testing powerful AI models without adequate controls. The lawsuit follows a Hugging Face intrusion that OpenAI’s own account traced to containment failures, missing monitors and mishandled warning signs.
Perplexity has launched Frontier Research to develop AI that can formulate questions, test hypotheses and produce interpretable discoveries. The company promises more than 10,000 GPUs, but existing automated-research systems show why computing resources and generated papers alone cannot establish scientific value.
The new lab led by Stanford professors Yejin Choi and Carlos Guestrin has backing from Nvidia, Cisco and Snowflake. It promises systems that improve on company work, but research benchmarks leave production costs and customer control unproven.
NVIDIA announced a Rubin HPC GPU with more than 50,000 CUDA cores, sparse data acceleration and seventh-generation NVLink. The design targets simulations used by scientists and AI agents, but missing performance, power and pricing details leave its practical advantage unproven.
OpenAI will open 50,000 new ChatGPT access grants on November 23, covering up to 150,000 researchers and collaborators. New recipients get Plus instead of Pro, with first-come selection replacing a lottery. Its pilot study reports perceived time savings but does not establish faster discovery.
OpenAI reports production healthcare workflows at CVS and Cigna and stronger computer-use evaluations. The results show progress across disconnected software, but leave complete-task reliability and total deployment costs unresolved.
President Donald Trump supported a return for Nvidia co-founder Jensen Huang, but the company has installed interim leadership and begun a successor search amid scrutiny of his conduct and alleged Chinese intelligence ties.
Mercedes-Benz plans California early access to city-driving and hands-free highway assistance in December, ahead of a 2027 U.S. rollout. Mixed road tests and unannounced subscription prices leave buyers waiting for evidence of everyday value.
Researchers estimate that 403 large U.S. data centers consumed 81.8 TWh in a year. Their regional-grid analysis reveals substantial fossil-fuel dependence, but cannot isolate AI demand or fully expose the underlying facility data.
HSBC lowered its Zhipu share-price target to HK$500 and delayed forecast adjusted breakeven by two years. Improving API margins and substantial fundraising cushion a business whose gross profit still covers little of its research bill.