Category: Knowledge & Reference · Draft date: 2026-11-22
The number everyone remembers from the Diamond Rapids spec sheet is "up to 256 P-cores." Read Intel's SC26 messaging together with the Hot Chips 2026 architecture details and the physical package Bull put on display, though, and the real bet of this Xeon generation sits on two other figures: 1.6TB/s of memory bandwidth and 1.28GB of last-level cache. Core count decides how much you can compute; bandwidth and cache decide how much of that compute you can actually feed. The recurring lesson of recent server CPU generations is that the latter is the binding constraint. As of this draft Intel has published no launch date. What follows is an evidence-chain teardown of the architecture, a comparison against shipping platforms, and acceptance advice for buyers.
After the P-core/E-core fork, which road does the flagship take?

Intel's server line has openly forked. The E-core density road belongs to Xeon 6+: 18A process, up to 288 E-cores, 12-channel DDR5, 96 lanes of PCIe Gen5, announced at Computex 2026 (StorageReview report, 2026-06-05) — it sells throughput per watt and consolidation density. The next step on the P-core performance road is Diamond Rapids: an 18A-P design with Foveros Direct 3D stacking and UCIe-S chiplet interconnects (Hot Chips 2026, StorageReview report 2026-08-24; Intel's official keynote handout, ID 926846, dated 2026-08-25). The instruction set adds APX, AVX 10.2 and enhanced AMX, and Intel calls it the "only x86 server CPU with built-in matrix compute" (vendor claim). 256 P-cores versus 288 E-cores is not one race: the former sells per-thread performance and matrix acceleration, the latter density and efficiency.
Physical evidence already exists. ServeTheHome reported on 2026-11-18 that Bull's Sequana XH31600 blade showed 4 Diamond Rapids packages at SC26 — 2 dual-socket nodes — on the LGA9324 socket (previously seen on ASRock's DMRD16-2L+ board at OCP Summit 2026), each package with 16-channel memory and blind-mate liquid cooling. Bull was the only vendor displaying real packages; the silicon is at the sampling stage. Tom's Hardware, at the same SC26, watched a Diamond Rapids server share a memory pool with a Xeon 6 system through a Montage MXC controller at the CXL booth — it ran on day one and was shut down the next day after noise complaints. Media estimate the earliest launch window at Q2 2027, with 2027-09 as the fallback if it slips (STH's speculation, not an official commitment).
The arithmetic behind 1.6TB/s: 16 channels × 12800 MT/s
Hot Chips 2026 states 16 memory channels at up to 12800 MT/s. Theoretical bandwidth = 16 × 12800 × 8 bytes = 1638.4 GB/s ≈ 1.6TB/s (calculated), matching the SC26 headline. The SC26 news piece did not say whether the figure is per socket; with 16 channels per CPU at Hot Chips, it should be a per-socket number (inferred). 12800 sits beyond mainstream RDIMM rates (6400) and maps to the MRDIMM Gen2 spec tier (inferred: Intel has not named the module type, but AMD's Venice SP7 already runs MRDIMM-12800 in StorageReview's 2026-11-10 review).
The per-core math is less flattering (calculated): 1638.4 ÷ 256 = 6.4 GB/s. The previous flagship, Granite Rapids (Xeon 6980P), delivers 12 × 8800 × 8 = 844.8 GB/s, or ÷ 128 = 6.6 GB/s per core. Total bandwidth rises +94% while cores rise +100% — per-core bandwidth actually dips. That is exactly why the 1.28GB cache has to exist: trading hit rate for bandwidth.
I/O moves a generation too: 128 lanes of PCIe Gen6 plus CXL 3.0 (Hot Chips figures). Against the 6980P's 96 lanes of Gen5, aggregate I/O bandwidth is roughly 2.67× (calculated: 128 × 64 GT/s ÷ 96 × 32 GT/s). And CXL 3.0 memory pooling is not a slide: the Montage demo at SC26 was a live machine sharing memory across generations.
Accounting for the 1.28GB cache hierarchy
1.28GB = 1280MB (converted), which spreads to 5MB per core across 256 cores (calculated). The 6980P offers 504MB ÷ 128 ≈ 3.9MB per core; the EPYC 9996 offers 1024MB ÷ 256 = 4MB per core. Mind the terminology: Intel disclosed a total "last-level cache" figure — the L2/LLC split, the topology, and whether any of it is 3D-stacked cache remain unpublished. Ask about the hierarchy at acceptance time; do not let LLC be marketed as L2. For pointer-chasing database workloads, a 5MB-per-core shared pool keeps more hot indexes on die. For virtualization, check whether LLC partitioning and QoS scale with the core count (details undisclosed).
Against shipping platforms
| Item | Diamond Rapids (announced) | Xeon 6980P (shipping) | EPYC 9996 (shipping) | Xeon 6+ (E-core) |
|---|---|---|---|---|
| Status | Sampling; media estimate Q2 2027 onward | Launched (Q3 2024) | GA 2026-11-10 | Announced at Computex 2026 |
| Process | 18A-P | Intel 3 | TSMC N2 | 18A |
| Max cores | 256 P-cores | 128 P-cores | 256 Zen 6c | 288 E-cores |
| Last-level cache | 1.28GB | 504MB | 1024MB (L3) | Not disclosed |
| Memory | 16 channels at 12800 MT/s | 12 channels MRDIMM-8800 | 16 channels MRDIMM-12800 | 12 channels DDR5 |
| Theoretical bandwidth (calculated) | 1638.4 GB/s | 844.8 GB/s | 1638.4 GB/s | — |
| PCIe | 128 lanes Gen6, CXL 3.0 | 96 lanes Gen5 | Gen6 | 96 lanes Gen5, CXL |
| TDP | Not disclosed | 500W | 600W | Not disclosed |
Note: theoretical bandwidth = channels × rate × 8 bytes (calculated). Diamond Rapids and Xeon 6+ are announced/sampling figures; measured data lives in the body text and sources.
What it means for HPC, databases and virtualization
HPC: bandwidth-hungry kernels (CFD, sparse matrices, STREAM-style codes) benefit directly. Venice's measured results already quantify the prescription: moving the same silicon from DDR5-8000 to MRDIMM-12800 lifted STREAM Triad bandwidth +58% and cut the large OpenFOAM solve time −21% (StorageReview review). Diamond Rapids writes the same prescription for the Intel platform, then layers AMX on top for CPU-side matrix acceleration.
Databases: a big LLC pushes down memory-access tail latency, but flat per-core bandwidth is the reminder that 256 cores do not equal 2× throughput. Budget memory-bound OLTP by per-core bandwidth, not by linear extrapolation from core count.
Virtualization: an all-P-core design removes big.LITTLE scheduling heterogeneity and consolidates densely; CXL 3.0 pooling gives memory tiering and overcommit a hardware path, with the SC26 demo as the prototype.
The heterogeneous sidekick: the GPU D270 shown alongside (formerly Crescent Island) carries 160GB of LPDDR5X at 1.5TB/s in a 350W air-cooled PCIe card, with partner configurations up to 480GB. Intel claims up to 5.9× the FP64 financial throughput of an RTX PRO 6000 Blackwell Server Edition (600W) — but StorageReview's caveats stand: pre-production silicon, vendor-run tests, and a financial comparison pitting 8 NVIDIA cards against 1 D270 with no statement on per-card normalization. Quote it; do not swallow it.
Acceptance advice for buyers and integrators
- No prepayments: there is no official launch date, only a media-estimated Q2 2027 start. Put the delivery window and breach liability in the contract.
- Archive measured bandwidth: require a per-socket STREAM Triad run and the ratio of measured to theoretical (1638.4 GB/s). For MRDIMM, confirm the QVL, 1DPC/2DPC topology and mixing rules.
- Pin down the cache hierarchy: the L2/LLC split behind 1.28GB, its topology and partition QoS. Accept against your own database or CFD workload, not just a SPEC score.
- Software-stack maturity: APX, AVX 10.2 and AMX need compiler and library support (oneAPI, GCC, MPI distributions). Ask for a porting checklist for your target workloads.
- Platform cost: the new LGA9324 socket means new boards; Bull already designs for blind-mate liquid cooling, so air-cooled rooms must budget for retrofits. In Pearl River Delta facilities constrained by power density, measure what today's platforms actually deliver before deciding to wait for 2027.
The Diamond Rapids story is not "more cores" — it is "better fed." With AMD already selling 16-channel MRDIMM-12800 off the shelf (GA 2026-11-10), whether this Xeon generation delivers its 1.6TB/s plus 1.28GB will be settled by measured data in 2027.
Sources (URL + date, all verified on 2026-11-22)
- StorageReview SC26 report (256 P-cores, 1.6TB/s, 1.28GB, D270 and vendor-comparison caveats): https://www.storagereview.com/news/intel-diamond-rapids-1-6tb-s-memory-bandwidth-1-28gb-cache-gpu-d270-sc26 , 2026-11-16
- StorageReview Hot Chips 2026 report (18A-P, 16 channels at 12800, 128 lanes Gen6, CXL 3.0): https://www.storagereview.com/news/intel-hot-chips-2026-256-core-diamond-rapids-crescent-island-with-480gb-for-inference-and-wildcat-lake-at-the-edge , 2026-08-24
- Intel official keynote handout (ID 926846; page accessible, body is a download, specs cross-checked via source 2): https://www.intel.com/content/www/us/en/content-details/926846/diamond-rapids-intel-s-next-generation-xeon-cpu-keynote-handout-for-customers.html , 2026-08-25
- ServeTheHome (Bull's physical package; body fetch limited, cross-checked via Zhiding syndication and search snippets): https://www.servethehome.com/bull-shows-intel-xeon-diamond-rapids-package-at-sc26/ , 2026-11-18
- Zhiding (Chinese syndication of ServeTheHome; LGA9324, 16 channels, launch-window speculation): https://www.zhiding.cn/server/2026/1119/3206513.shtml , 2026-11-19
- Tom's Hardware (live CXL-booth demo; page returned title only, date per index timestamp, content cross-checked via 3dnews report of 2026-11-20): https://www.tomshardware.com/pc-components/cpus/intel-diamond-rapids-shows-signs-of-life-in-sc26-demo-running-live-server-next-gen-xeon-7-cpu-breaks-cover-at-partner-booths-no-showcase-from-intel , 2026-11-18
- Intel ARK Xeon 6980P spec page (fetch returned navigation only; specs verified via search snippet of the official page): https://www.intel.com/content/www/us/en/products/sku/240777/intel-xeon-6980p-processor-504m-cache-2-00-ghz/specifications.html , launched Q3 2024
- StorageReview EPYC Venice SP7 review (9996 specs, MRDIMM-12800 measurements, GA date): https://www.storagereview.com/review/amd-epyc-venice-sp7-review , 2026-11-10
- StorageReview Xeon 6+ Computex report (288 E-cores, 18A, 12-channel DDR5, 96 lanes Gen5): https://www.storagereview.com/news/intel-launches-xeon-6-on-18a-with-288-e-cores-e835-200gbe-ethernet-and-crescent-island-gpu-details , 2026-06-05
