On 2025-12-08, the PC Cluster Consortium held PCCC25 — its 25th-anniversary symposium, themed "The HPC Frontier" (HPC 最前线) — in Tokyo. Intel senior solution architect Takafuji (髙藤良史) opened with "Advancing the Frontier," a 22-slide deck. The deck has received little public analysis, and Chinese-language coverage is especially scarce, but slide 8's HPC performance chart is one of the few public documents that compares DDR5-6400 and MRDIMM-8800 workload by workload on the same CPU. This article takes that chart as its axis, cross-checks it against Intel's official documentation and third-party measurements, and finishes the math the slides leave open.
The Bandwidth Math: Per-Core Rations Stuck at 4.8 GB/s

A memory-bound workload is, at bottom, compute waiting for data. Run the per-core bandwidth calculation from public specifications (all values below are calculated). The AMD EPYC 9654 has 12 channels of DDR5-4800 and an official per-socket bandwidth of 460.8 GB/s; divided by 96 cores, that is 4.8 GB/s per core. The Xeon 6980P with DDR5-6400 has a theoretical bandwidth of 12 channels × 8 bytes × 6400 MT/s = 614.4 GB/s; divided by 128 cores — still 4.8 GB/s. Without a change in memory form factor, the flagship CPU's per-core ration does not move across generations: core counts inflate, but the bandwidth each core can claim does not. That is the memory wall.
Large caches soften part of the problem: the 6980P's 504 MB of L3 (slide 6) absorbs a substantial share of repeated accesses. But sparse matrices and stencil kernels stream data through once and never look back — no cache can rescue those; only bandwidth can. HPCG is MRDIMM's best advertisement precisely because its irregular access pattern collects almost no cache dividend.
Switch to MRDIMM-8800 and the theoretical figure becomes 844.8 GB/s, or 6.6 GB/s per core — a 37.5% increase. That is the entire motivation for MRDIMM: not faster chips, a wider road.
"2× Transfer" Is Not 8800 ÷ 6400: How to Read the Measured Scaling
Slide 7 draws the MRDIMM structure: Rank-0 and Rank-1 on the module, a multiplexing chip at the bottom, labeled "2× DDR5 transfer." Many readers will take that to mean "8800 is twice 6400." It does not. The 2× refers to multiplexing two ranks inside the module: the DRAM devices actually run at 4400 MT/s, the mux chip interleaves the two ranks' data streams, and the host side sees 8800 MT/s. (Micron's FMS25 talk, as relayed in Chinese-language coverage, explains it this way; Intel's official November 2024 newsroom release phrases it as letting data flow through both ranks simultaneously.) Against DDR5-6400 RDIMM, the interface rate is really 1.375× — consistent with Intel China's product brief claim of "more than 37% higher bandwidth," whose footnote explicitly names DDR5-6400 RDIMM as the comparison baseline.
Slide 8's measured chart normalizes the EPYC 9654 to 1.0; higher is better:
| Workload | DDR5-6400 | MRDIMM-8800 | Relative gain (calculated) |
|---|---|---|---|
| HPCG | 1.52 | 2.01 | +32% |
| Stream Triad | 1.31 | 1.70 | +30% |
| OpenFOAM (geomean of 2) | 1.71 | 2.05 | +20% |
| MPAS-A (geomean of 2) | 1.55 | 1.84 | +19% |
| NEMO (geomean of 2) | 1.46 | 1.75 | +20% |
| Ansys CFX (geomean of 2) | 1.46 | 1.69 | +16% |
| Overall geomean, 15 workloads | 1.51 | 1.67 | +11% |
Read the normalization carefully: the 6980P on DDR5-6400 already posts an overall geomean of 1.51 — cores and cache alone put it 50% ahead of the 9654 — and MRDIMM adds roughly 11% on top. The chart spans five categories (the slides' own taxonomy): industry benchmarks, earth-system modeling, finance, life and materials science, and manufacturing. Intel is telling an industry-wide story, not a single-benchmark one.
Third-party cross-check: in Phoronix's dual-socket 6980P testing (OpenBenchmarking result 2410037, September–October 2024), HPCG at a 104³ grid rose from 130.84 to 170.95 GFLOP/s — a gain of about 30%, directionally consistent with the slides. But in that same dual-socket run, Stream Triad measured about 893 GB/s on DDR5 versus about 880 GB/s on MRDIMM — essentially flat — while Scale gained 12%. The overall geomean was only 1.06×, with MRDIMM winning 77% of the individual tests. The gap between the vendor chart and third-party results is not fraud; it is configuration. Single versus dual socket, SNC partitioning, and NUMA binding all change STREAM behavior dramatically — which is exactly what acceptance testing should watch.
AMX Plus Bandwidth: The Two Legs of HPC/AI
The 6980P's other selling point is AMX, the matrix extensions, with FP16 support added in this P-core generation (slide 6). The product brief gives concrete throughput: 2048 floating-point operations per core per cycle at INT8, 1024 at BF16/FP16, and matrix multiply-accumulate up to 16× faster than AVX-512. But a matrix engine's appetite scales with its compute; starve it of bandwidth and AMX idles. Intel's official claims — "up to 2× BERT-Large training and inference versus fifth-generation Xeon" and "up to 2.3× HPCG with MRDIMM" — are combined compute-plus-bandwidth results, not the credit of any single feature. The practical meaning of FP16 support: models trained in FP16 can run CPU inference without a precision conversion. Slide 10 lists the oneAPI libraries (oneMKL, oneDNN) and tools (VTune, Advisor) alongside — AMX gains are realized through libraries and compilers, so acceptance testing must use the target software stack, not raw instruction peaks.
The back half of the deck points at agentic AI: orchestration, data processing, and tool calls land on the CPU; prefill and decode land on the GPU. CPU-side inference and data preprocessing are exactly the "mid-size models plus high bandwidth" workload where MRDIMM plus AMX beats chasing HBM on cost. For Pearl River Delta teams building head nodes for GPU clusters or doing CPU inference, that is a pragmatic combination.
Who Gains, Who Stays Put
The same chart has a cold side: HPL 1.78 → 1.75, GROMACS 1.30 → 1.29, LAMMPS 1.70 → 1.71, RELION 2.01 → 1.96, and the FSI kernel 1.72 → 1.71. Compute-bound or cache-friendly workloads collect almost none of the MRDIMM dividend, and a few slip slightly — the added latency of the multiplexing buffer chip is one likely cause (the author's inference). WRF (+9%), Fluent (+7%), and LS-DYNA (+3%) sit in the middle.
Two conclusions follow. First, a 15-workload geomean of +11% does not mean your application gets +11%; the gain depends on the memory intensity of your kernel mix. Second, run your own application benchmarks before paying the MRDIMM premium — for HPL-class workloads, the money is better spent on more cores or a larger L3. In practice, profile the memory-bound fraction with VTune or Advisor first: code that sits at the bandwidth ceiling with low cache hit rates is MRDIMM's target customer; code pinned at the compute ceiling gains nothing from new memory.
Configuration and Acceptance: A Checklist for Pearl River Delta Procurement
On configuration: confirm the SKU supports MRDIMM at all (Fujitsu's PRIMERGY catalog explicitly warns that some CPUs do not). Populate all 12 channels, one DIMM per channel (1DPC); do not leave channels empty to save money — Phoronix ran a dedicated 6/8/10/12-DIMM scaling test in February 2026 to warn about exactly this. Then there is power: Micron's FMS25 talk (as relayed by the Tencent Cloud developer community) shows per-DIMM power rising from about 11 W for RDIMM to 18–21 W (21.2 W at 256 GB). Twelve DIMMs put roughly 240 W into memory alone (calculated); rack power and airflow need to be recalculated, and 18 W is a thermal-design watershed.
On acceptance: use Intel MLC in pure-read mode for the peak (the same relayed FMS25 figures: about 743 GB/s on a single socket with 12 channels of MRDIMM-8800, about a third above DDR5-6400's ~560 GB/s), then use STREAM and your own applications for the real gain. Enable SNC and bind NUMA, or you will be measuring cross-die bandwidth — in that matrix, local reads run about 274 GB/s versus about 212 GB/s across dies. Do not treat a vendor's normalized chart as an absolute promise; the slides themselves note that results may vary.
On procurement timing: memory prices are swinging hard in 2026, and MRDIMM supply is concentrated in a handful of original manufacturers. University labs, CFD-driven manufacturers, and meteorology/oceanography teams in Guangzhou and the Pearl River Delta should pilot one or two nodes on their own workloads before sizing a cluster. Write "MRDIMM-8800, all 12 channels populated, acceptance bandwidth threshold" into the contract — not just "Xeon 6 server." Plan capacity and expansion separately: MRDIMM per-DIMM capacities span 32–256 GB (Micron's materials), so bandwidth and capacity can be bought in one pass; if the bottleneck is capacity rather than bandwidth, Xeon 6's CXL 2.0 memory expansion (per the product brief: up to 64 lanes at 32 GT/s) is a cheaper path. The two are complementary, not exclusive.
Sources
- PC Cluster Consortium: PCCC25 "The HPC Frontier" (HPC 最前线) program and presentation materials (髙藤良史 [Takafuji], "Advancing the Frontier," talk 2025-12-08; materials published 2025-12-26). http://www.pccluster.org/ja/event/pccc25/program/ ; slides: http://www.pccluster.org/ja/event/pccc25/data/PCCC25_8-1_intel_takafuji.pdf (full text downloaded and parsed; the slide 8 chart visually verified against a rendering)
- Intel: "英特尔至强 6 处理器家族产品简介" (Intel Xeon 6 Processor Family Product Brief, Chinese edition), 2025-02. https://www.intel.cn/content/dam/www/central-libraries/cn/zh/documents/2025-05/25-dcai-the-xeon-6-processor-family-product-brief-update.pdf
- Intel Newsroom: "New Ultrafast Memory Boosts Intel Data Center Chips," 2024-11-15. https://www.intel.com/content/www/us/en/newsroom/news/new-ultrafast-memory-boosts-intel-xeon-chips.html
- OpenBenchmarking/Phoronix: "Granite Rapids MRDIMM vs. DDR5 Benchmarks," 2024-09-30/10-03. https://mail.openbenchmarking.org/result/2410037-NE-2410039NE61
- AMD: EPYC 9654 product page (per-socket bandwidth 460.8 GB/s). https://www.amd.com/ja/products/processors/server/epyc/4th-generation-9004-and-8004-series/amd-epyc-9654.html
- Micron: FMS25, "Beyond bandwidth: MRDIMMs' edge in the data center revolution" (Potter, 2025-08-07). https://files.futurememorystorage.com/proceedings/2025/20250807_DRAM-303-1_Potter.pdf (original PDF fetch failed; the 743 GB/s and 18–21 W figures are relayed via a Tencent Cloud developer community analysis, 2026-01-13): https://cloud.tencent.com.cn/developer/article/2616612
- Fujitsu: PRIMERGY RX2530 M8 PC cluster catalog, 2026-02. https://www.fsastech.com/ja-jp/products/primergy/pccluster/pdf/pcc_ansys-ls-dyna_performance_m8gnr.pdf
- Phoronix: "Intel Xeon 6 Granite Rapids Memory Scaling Performance From 6 To 12 MRDIMMs," 2026-02-17 (site blocks crawlers; verified via search-engine index summary). https://www.phoronix.com/review/intel-xeon-mrdimm-scaling
Note: values marked "calculated" are the author's computations from the public specifications and chart values above, including the table's relative-gain column and the percentages converted from the chart in the body text (+12%, +9%, +7%, +3%, etc.). The slide chart shows EPYC 9654-normalized relative values, not absolute bandwidth.
