UR-CP400: How to Read a 22/GHz RISC-V Core Claim
Every RISC-V performance claim eventually runs into the same wall: a big number with no denominator. “3 GHz”, “70 points”, “8-wide”, “双倍于上一代” — without a clock, a baseline version, and a statement of whether the silicon exists, none of it is comparable to anything. Which is why the most useful thing UltraRISC (超睿科技, Shanghai) published at RISC-V Summit China 2026 (Shenzhen, 18–20 October) was not its headline number but its denominator: SPECint_base2006 above 22 per GHz for the fourth-generation UR-CP400 core — a figure the company itself pins to an FPGA prototype, not to silicon.
That single qualifier is the difference between a benchmark and a marketing slide. This article walks through what was actually disclosed, what the generational numbers say when you normalise them, why per-GHz is the only honest way to compare cores at this stage — and the three PPA data points that are still missing and that any licensing discussion will demand.
What UltraRISC Actually Showed
The venue matters as much as the number. UR-CP400 was presented by CTO Jiang Jiang (蒋江) in the data-centre and HPC track of the summit, alongside an FPGA prototype system running on the booth floor. The distinction the company was careful to make — and which most Chinese-language coverage blurred — is that UR-CP400 is a processor core IP, not a chip. What ships to a customer is a design, and what a licensee integrates determines the final clock, power and memory subsystem.
Alongside the core, UltraRISC showed system-level products built on its existing silicon: UR-DP1000, an eight-core desktop-class CPU with 8 × UR-CP100 cores on a 12 nm process, plus a water-cooled “all-domestic AI edge appliance”, a 64-core multi-node cluster, and a cross-host “many-virtual-machines-as-one” demo. That split — fast core as a future IP, modest core as shipping silicon — is the honest state of the art for every RISC-V core vendor in this performance band.
The Roadmap in Numbers
UltraRISC maintains three core families: UR-CP for high performance, UR-CR for high reliability and UR-CE for high efficiency. The CP line is the one with published microarchitectural parameters, and its five generations can be laid out in a single table — which is rarer than it should be in this industry.
| Generation | Issue / pipeline | ISA | Performance claim | Status |
|---|---|---|---|---|
| UR-CP100 | Out-of-order, 4-issue, 12-stage | RV64GCBHX (X = UltraRISC high-perf ALU/LS extension), RVH 1.0 | SPECint2006 10.4/GHz, SPECfp2006 12.0/GHz | Silicon: 12 nm, 2.0–2.3 GHz, shipping in UR-DP1000 |
| UR-CP200 / CP300 | Out-of-order, 6-issue, improved branch prediction and memory access | RVA23-class | Vendor positions them against Arm Cortex-A78 and Neoverse N2 respectively (no published per-GHz figure) | IP deliverable (per vendor) |
| UR-CP400 | Out-of-order, 8-decode / 8-issue, 14-stage, ROB 750+ (configurable) | RVA23 compliant, RVV 1.0 | SPECint_base2006 > 22/GHz (FPGA prototype); > 3 GHz at 5 nm (evaluation) | FPGA prototype running; IP delivery planned Q1 2027 |
| UR-CP500 | Out-of-order, 10-issue | RVA23 + RVV 1.0, configurable VPU, in-cluster matrix acceleration | Planned > 26/GHz | Roadmap; Q1 2028 delivery target |
Two things in that table are worth pausing on. First, the jump from 4-issue to 8-issue in one generation, without an intermediate 8-wide part: UR-CP100’s 4-issue front end and 750-entry reorder buffer belong to different eras of core design. Second, the company’s own framing of the trajectory — each generation delivering “on average more than 20 % performance improvement over the previous one” — is arithmetically inconsistent with the two published endpoints: 10.4/GHz to >22/GHz is more than 2 × per clock. The gap is explained by the fact that the mid-generations only have relative claims (A78, N2) and no numeric per-GHz figure.
Execution Resources: Where the 22 Comes From
The published execution width is the most concrete thing UltraRISC disclosed, and it is specific enough to be checked against the number:
| Resource | UR-CP400 (as disclosed) |
|---|---|
| Integer units | 6 × ALU |
| Floating-point units | 4 × FPU |
| Branch units | 2 × BU |
| Vector units | 1–2 × VPU, count and data width configurable; VLEN default 256 bit |
| Memory units | 2 full Load/Store + 2 Load = 4 ports, shared between scalar and vector |
| Cache hierarchy | L1 and L2 per-core private; L3 shared across cores in a cluster |
| Matrix support | Cluster-shared RISC-V standard matrix unit with direct L3 access |
| System interfaces | IOMMU, AIA, RERI server companion IP |
| Virtualization | Hardware-assisted, software/hardware co-designed nested virtualization |
Six ALUs, four FPUs, two branch ports and up to two 256-bit vector units behind a 750-entry window is a genuinely wide machine — wider in issue width than XuanTie C950’s eight decode with a 1000+ entry window, and in the same architectural class. The four shared memory ports are the part that determines whether that width can be fed; it is also, notably, fewer load/store ports per issue slot than several x86 designs field, which is exactly where a per-GHz integer number tends to be bounded.
Why Per-GHz, and Where It Stops Being Honest
Dividing a SPEC score by clock frequency isolates microarchitectural efficiency from process and physical design. It is the right metric for comparing cores — and the wrong metric for comparing products. Normalising the published figures gives a clean picture of the generational step:
| Core | Claimed integer performance | Normalised | Evidence level |
|---|---|---|---|
| UR-CP100 (in UR-DP1000) | SPECint2006 10.4/GHz at 2.0–2.3 GHz | 10.4/GHz | Silicon, 12 nm, shipping |
| UR-CP400 | SPECint_base2006 > 22/GHz | > 22/GHz | FPGA prototype |
| XuanTie C950 (for reference) | SPECint2006 > 70 points at 3.2 GHz | ≈ 21.9/GHz by arithmetic | Vendor claim, 2026-11-02 award coverage |
Read together, those three rows say something more interesting than either company’s press release: two independent Chinese teams have converged on roughly 22 SPECint2006 points per gigahertz as the target for an RVA23-class, 8-wide, out-of-order core. That is the first time a per-clock ceiling has become a shared reference point in the domestic RISC-V ecosystem — and, if both numbers survive silicon, more than double UR-CP100’s 10.4/GHz, which puts the class in range of where Arm’s performance cores sat a couple of generations back rather than a decade back.
① FPGA, not silicon. An FPGA prototype runs at a small fraction of target clock and its block RAM and DSP timing are not representative. Per-GHz is designed to cancel that out, but the cancellation is imperfect: memory-level parallelism, cache latency in cycles and clock-relative I/O all change in real silicon.
② SPECint2006, not 2017. The two suites are not interchangeable and no conversion factor is legitimate. Every vendor still quoting 2006 is doing so because that is what their older published numbers use — comparisons against a 2017-based figure are meaningless.
③ RVA23 “compliant” is a design statement. Compliance is established by running the RISC-V Architecture Test suite plus the profile’s required extension tests, and no third-party result has been published for UR-CP400.
The PPA That Is Still Missing
Core IP is bought on three numbers — performance, power, area — and only one of the three was disclosed. UltraRISC published no power figure, no core area in mm², and no sustained clock in a specific process corner (“> 3 GHz at 5 nm” is an evaluation estimate, and 5 nm is not necessarily the node a licensee will use). It also did not publish licensing terms or pricing.
For context on how much those missing numbers matter, the shipping part is fully documented: UR-DP1000 pairs eight UR-CP100 cores in two clusters of four, each cluster sharing 4 MB L3 with a 16 MB LLC across both, 64 KB L1I + 64 KB L1D + 512 KB L2 per core, dual-channel DDR4-3200 with ECC up to 128 GB, 24 lanes of PCIe 4.0, FCBGA-1155 at 35 × 35 mm, 12.8 × 12.8 mm die, 30 W TDP, taped out March 2024, with KVM overhead reported at roughly 3 %. That is a complete, checkable dataset. UR-CP400 has a microarchitecture diagram and one benchmark line.
The Memory Wall Hiding Behind a Per-Clock Number
Per-GHz figures are measured with a benchmark whose working set is chosen to fit the cache hierarchy. Inside a cluster that hierarchy is L1 and L2 per core plus a shared L3, with the matrix unit reading L3 directly — a sensible arrangement for tile kernels. Outside the cluster, though, the shipping part tells you what platform class these cores are aimed at: UR-DP1000 carries dual-channel DDR4-3200. Two 64-bit channels at 3200 MT/s is 51.2 GB/s of peak theoretical bandwidth for eight cores — roughly 6.4 GB/s per core. A single SpacemiT V100, for comparison, is specified with eight channels of DDR5. That gap is not a criticism of the core; it is the reason a core IP announcement cannot be read as a server announcement. Whoever integrates UR-CP400 decides whether the 22/GHz survives contact with a real memory subsystem, and the L3 sharing ratio, NoC topology and DIMM count will move the delivered number more than any microarchitectural parameter in the table above.
Four Things to Watch Before Q1 2027
- A silicon per-GHz number, not an FPGA one. If UR-CP400 lands near 22/GHz in a test chip, the generation-over-generation claim holds; if it lands nearer 16–18/GHz, the FPGA extrapolation was optimistic and the roadmap’s “20 % per generation” line was doing more work than the microarchitecture.
- Whether RVA23 compliance is tested in public. The architecture tests are open; publishing a run log costs nothing and settles the question that every OS distributor will ask.
- Where the vector units land in the default configuration. One 256-bit VPU versus two is roughly a doubling of vector throughput and a large area difference. Which one a licensee gets by default determines whether UR-CP400 competes with C950 on AI-adjacent workloads or only on scalar integer.
- The first licensee and the node. A core announced at 5 nm that first appears in a 12 nm product tells you the economics the market will actually bear.
The Takeaway
UR-CP400 is the clearest statement yet that Chinese RISC-V core design has stopped treating integer IPC as a second-order problem. Eight-wide decode, a 750-entry window, configurable 256-bit vector, IOMMU/AIA/RERI hooks and nested-virtualization acceleration is a server-class feature list, and the 22/GHz figure — cautiously read — puts it in the same per-clock band as XuanTie’s flagship. What it does not yet have is the three numbers a customer signs against: silicon per-GHz, power, and area. Until those exist, the right way to cite UR-CP400 is exactly the way UltraRISC itself did: “above 22 per gigahertz, FPGA configuration, IP delivery planned Q1 2027”. That sentence is worth more than a press release number, because it can be checked.
中文摘要
在 2026 RISC-V 中国峰会(10 月 18–20 日,深圳)上,超睿科技(UltraRISC)由执行总裁兼 CTO 蒋江发布了第四代高性能处理器核 UR-CP400,并在展台演示了 FPGA 原型系统。关键事实需要先分清:这是处理器核 IP,不是芯片;计划 2027 年 Q1 交付设计文件,而非可上电的硅片。
已披露规格:符合 RVA23 规范、支持 RVV 1.0;8 译码 / 乱序 8 发射 / 14 级流水线,重排序缓冲区 750+ 项(可配置);执行资源为 6×ALU + 4×FPU + 2×BU + 1~2×VPU(向量长度可配置,缺省 256 bit);访存为 2 个全功能 Load/Store + 2 个 Load,标量与向量共用 4 个访存单元;L1/L2 核私有、L3 簇内共享,簇内另有符合 RISC-V 标准的矩阵单元直接访问 L3;可对接 IOMMU / AIA / RERI 服务器配套 IP,虚拟化侧提供硬件加速的软硬件协同嵌套虚拟化,并针对 Agentic AI 的低延迟、高并发混合负载做了微架构优化。
性能口径(务必带限定词):FPGA 仿真配置下 SPECint_base2006 超过 22/GHz;5nm 工艺评估下主频大于 3GHz。前者是FPGA 而非硅数据,基准是 2006 版而非 2017 版,且功耗、面积、量产频率一概未公布——而这三项恰恰是 IP 授权谈判的硬指标。用绝对分数来看:上一代 UR-CP100(已量产于 8 核桌面 CPU UR-DP1000,12nm、2.0–2.3GHz、30W)为 10.4/GHz,故 CP400 约为其 2.1 倍的每时钟整数性能。
一个可交叉验证的巧合:达摩院玄铁 C950 公布的口径为 3.2GHz 下 SPECint2006 单核首破 70 分,折算约 21.9/GHz,与 UR-CP400 的 >22/GHz 基本同档。这意味着国内两家独立团队已在 约 22 分/GHz 这一「每时钟天花板」上形成事实共识——这是 RVA23 时代第一次出现可互相比对的统一坐标。
路线图:UR-CP100(4 发射/12 级,RVA22 级,已硅验证)→ UR-CP200/CP300(6 发射,厂商称分别对标 Cortex-A78 与 Neoverse N2,无可比每 GHz 数字,IP 可交付)→ UR-CP400(8 发射,2027Q1)→ UR-CP500(10 发射、簇内矩阵加速,规划 >26/GHz,2028Q1)。厂商称「每代平均提升 20% 以上」,但与 10.4→22 的两个端点数字并不严格自洽——中间两代只有相对表述,没有数字,这正是缺口所在。
结论:UR-CP400 的微架构与系统配套(IOMMU/AIA/RERI、嵌套虚拟化)已经是一张服务器级核的清单,但它仍缺客户签约时要看的三件事:硅上的每 GHz 成绩、功耗、面积。在此之前,正确的引用方式就是厂商自己的表述——「>22/GHz,FPGA 配置,IP 计划 2027Q1 交付」。
Кратко (RU)
На саммите RISC-V Summit China 2026 (Шэньчжэнь, 18–20 октября) компания UltraRISC (Шанхай) показала ядро четвёртого поколения UR-CP400. Важное уточнение: это IP-ядро, а не чип; поставка проекта запланирована на I квартал 2027 года.
Заявленные характеристики: соответствие RVA23 и поддержка RVV 1.0; 8 декодирований / 8 выдач вне очереди / 14 стадий, буфер переупорядочивания 750+ (настраиваемый); 6 ALU + 4 FPU + 2 BU + 1–2 VPU (VLEN по умолчанию 256 бит); четыре порта памяти (2 полных Load/Store + 2 Load), общих для скалярных и векторных операций; L1/L2 приватные, L3 общий для кластера, плюс кластерный матричный блок с прямым доступом к L3; интерфейсы IOMMU / AIA / RERI и аппаратно-ускоренная вложенная виртуализация.
Метрика с оговорками: SPECint_base2006 выше 22 на ГГц — но на FPGA-прототипе, а не на кремнии, и по базе 2006, а не 2017 года. Тактовая частота, мощность и площадь не раскрыты. Предыдущее ядро UR-CP100 в серийном UR-DP1000 (12 нм, 2,0–2,3 ГГц, 30 Вт) даёт 10,4/ГГц, то есть прирост примерно в 2,1 раза на такт.
Перекрёстная проверка: XuanTie C950 заявляет более 70 баллов SPECint2006 при 3,2 ГГц — это около 21,9/ГГц, то есть тот же диапазон. Похоже, отрасль сходится к ~22 баллам на ГГц как ориентиру для 8-широкого ядра класса RVA23.
Честность: без данных по кремнию, мощности и площади цифра 22/ГГц позволяет судить о направлении архитектуры, но не окупает решение о лицензировании. Дорожная карта: CP200/CP300 (6 выдач, ориентиры A78/N2), CP500 (10 выдач, >26/ГГц, 2028 Q1).
Resumen (ES)
En RISC-V Summit China 2026 (Shenzhen, 18–20 de octubre), UltraRISC (Shanghái) presentó el núcleo de cuarta generación UR-CP400. Conviene precisarlo: es IP de núcleo, no un chip; la entrega del diseño está prevista para el primer trimestre de 2027.
Especificaciones divulgadas: conforme a RVA23 y con RVV 1.0; 8 decodificaciones / 8 emisiones fuera de orden / 14 etapas con ROB de 750+ entradas configurables; 6 ALU + 4 FPU + 2 BU + 1–2 VPU (VLEN por defecto 256 bits); cuatro puertos de memoria (2 Load/Store completos + 2 Load) compartidos entre escalar y vectorial; L1/L2 privadas por núcleo, L3 compartida por clúster, más una unidad matricial de clúster con acceso directo a L3; IOMMU / AIA / RERI y virtualización anidada asistida por hardware.
La métrica, con sus límites: SPECint_base2006 por encima de 22 por GHz, pero sobre un prototipo FPGA, no sobre silicio, y con la base 2006, no 2017. No se publicaron frecuencia final, potencia ni área. El núcleo anterior UR-CP100, ya en producción en el UR-DP1000 (12 nm, 2,0–2,3 GHz, 30 W), rinde 10,4/GHz: el salto es de aproximadamente 2,1× por ciclo.
Contraste útil: XuanTie C950 declara más de 70 puntos SPECint2006 a 3,2 GHz, es decir unos 21,9/GHz. Dos equipos independientes parecen converger hacia ~22 puntos/GHz como referencia para un núcleo de 8 vías clase RVA23.
Conclusión honesta: la lista de características es de clase servidor, pero faltan los tres datos que un cliente exige antes de firmar: rendimiento en silicio por GHz, potencia y área.
Résumé (FR)
Au RISC-V Summit China 2026 (Shenzhen, 18–20 octobre), UltraRISC (Shanghai) a présenté son cœur de quatrième génération UR-CP400. Précisons d’emblée : il s’agit d’IP de cœur, pas d’une puce ; la livraison du design est prévue au premier trimestre 2027.
Caractéristiques publiées : conforme RVA23, prend en charge RVV 1.0 ; 8 décodages / 8 émissions hors-ordre / 14 étages, ROB de 750+ entrées configurables ; 6 ALU + 4 FPU + 2 BU + 1–2 VPU (VLEN par défaut 256 bits) ; quatre ports mémoire (2 Load/Store complets + 2 Load) partagés entre scalaire et vectoriel ; L1/L2 privés par cœur, L3 partagé au niveau du cluster, plus une unité matricielle de cluster accédant directement à la L3 ; interfaces IOMMU / AIA / RERI et virtualisation imbriquée assistée par le matériel.
La mesure, avec ses limites : SPECint_base2006 supérieur à 22 par GHz, mais sur prototype FPGA et sur la base 2006, non 2017. Ni fréquence finale, ni consommation, ni surface ne sont publiées. Le cœur précédent UR-CP100, déjà en production dans l’UR-DP1000 (12 nm, 2,0–2,3 GHz, 30 W), atteint 10,4/GHz : le gain est d’environ 2,1× par cycle.
Recoupement : XuanTie C950 revendique plus de 70 points SPECint2006 à 3,2 GHz, soit environ 21,9/GHz. Deux équipes indépendantes semblent converger vers ~22 points/GHz comme référence pour un cœur 8 voies de classe RVA23.
Lecture honnête : la fiche est de classe serveur, mais les trois chiffres décisifs — performance silicium par GHz, consommation, surface — restent absents.
Kurzfassung (DE)
Auf dem RISC-V Summit China 2026 (Shenzhen, 18.–20. Oktober) stellte UltraRISC (Shanghai) seinen Kern der vierten Generation UR-CP400 vor. Wichtig: Es handelt sich um Kern-IP, nicht um einen Chip; die Auslieferung des Designs ist für Q1 2027 geplant.
Veröffentlichte Daten: RVA23-konform, RVV 1.0; 8 Dekodierungen / 8 Ausgaben out-of-order / 14 Stufen, 750+ konfigurierbare ROB-Einträge; 6 ALU + 4 FPU + 2 BU + 1–2 VPU (VLEN standardmäßig 256 Bit); vier Speicherports (2 volle Load/Store + 2 Load), gemeinsam für Skalar und Vektor; L1/L2 pro Kern privat, L3 clusterweit geteilt, dazu eine Cluster-Matrixeinheit mit direktem L3-Zugriff; IOMMU / AIA / RERI sowie hardwarebeschleunigte verschachtelte Virtualisierung.
Die Kennzahl mit Einschränkungen: SPECint_base2006 über 22 pro GHz — jedoch auf einem FPGA-Prototyp, nicht auf Silizium, und gegen die 2006er Basis, nicht 2017. Takt, Leistungsaufnahme und Fläche bleiben offen. Der Vorgängerkern UR-CP100 im bereits ausgelieferten UR-DP1000 (12 nm, 2,0–2,3 GHz, 30 W) erreicht 10,4/GHz — also rund 2,1× pro Takt.
Quervergleich: XuanTie C950 nennt über 70 Punkte SPECint2006 bei 3,2 GHz, also etwa 21,9/GHz. Zwei unabhängige Teams scheinen sich auf ~22 Punkte/GHz als Referenz für einen 8-fach-Kern der RVA23-Klasse zuzubewegen.
Fazit: Die Merkmalsliste ist Server-Klasse; es fehlen die drei Zahlen, gegen die Kunden unterschreiben — Siliziumleistung pro GHz, Leistungsaufnahme, Fläche.
خلاصه (FA)
در RISC-V Summit China 2026 (شنژن، ۱۸ تا ۲۰ اکتبر) شرکت UltraRISC (شانگهای) هسته نسل چهارم UR-CP400 را معرفی کرد. نکته مهم: این یک IP هسته است، نه یک تراشه؛ تحویل طراحی برای سهماهه اول ۲۰۲۷ برنامهریزی شده است.
مشخصات منتشرشده: سازگار با RVA23 و پشتیبانی از RVV 1.0؛ ۸ رمزگشایی / ۸ ارسال خارج از ترتیب / ۱۴ مرحله خط لوله با بافر بازآرایی بیش از ۷۵۰ ورودی قابل تنظیم؛ ۶ واحد ALU + ۴ واحد FPU + ۲ واحد انشعاب + ۱ تا ۲ واحد برداری (طول بردار پیشفرض ۲۵۶ بیت)؛ چهار پورت حافظه مشترک بین اسکالر و برداری؛ L1/L2 خصوصی هر هسته، L3 مشترک در خوشه و یک واحد ماتریسی خوشهای با دسترسی مستقیم به L3؛ رابطهای IOMMU / AIA / RERI و مجازیسازی تودرتوی شتابیافته با سختافزار.
معیار با قیدهایش: SPECint_base2006 بیش از ۲۲ بر گیگاهرتز — اما روی نمونه FPGA نه سیلیکون، و بر پایه ۲۰۰۶ نه ۲۰۱۷. فرکانس نهایی، توان و مساحت اعلام نشدهاند. هسته پیشین UR-CP100 در UR-DP1000 (۱۲ نانومتر، ۲٫۰ تا ۲٫۳ گیگاهرتز، ۳۰ وات) عدد ۱۰٫۴ بر گیگاهرتز دارد؛ یعنی جهش حدود ۲٫۱ برابر در هر سیکل.
مقایسه: XuanTie C950 بیش از ۷۰ امتیاز SPECint2006 در ۳٫۲ گیگاهرتز را اعلام کرده که حدود ۲۱٫۹ بر گیگاهرتز است. به نظر میرسد دو تیم مستقل به حدود ۲۲ امتیاز بر گیگاهرتز بهعنوان مرجع هسته هشتمسیره کلاس RVA23 نزدیک شدهاند.
نتیجه صادقانه: فهرست ویژگیها در سطح سرور است، اما سه عددی که مشتری بر آن قرارداد میبندد — کارایی سیلیکون بر گیگاهرتز، توان و مساحت — هنوز منتشر نشده است.
Sources / 参考来源
- UltraRISC summit recap (2026-11-02) — UR-CP400 execution resources, cache hierarchy, IOMMU/AIA/RERI, nested virtualization, CP500 roadmap — primary technical source
- UltraRISC developer blog — UR-DP1000 chip specification (UR-CP100 4-issue/12-stage, 10.4/GHz SPECint2006, 12.0/GHz SPECfp2006, 12 nm 2.0–2.3 GHz, 30 W, cache and I/O)
- EEFocus — UR-DP1000 tape-out March 2024, die 12.8 × 12.8 mm, KVM overhead ~3 %, NoC with cache coherence
- Pandaily — UR-CP400 as core IP, >22/GHz on FPGA, delivery Q1 2027, UR-CP100 comparison at 10.4/GHz
- Analysis (Chinese) — why SPECint2006/GHz cannot support a licensing decision without PPA; pricing and licensing terms not disclosed
- Coverage of the UR-CP100 → CP500 roadmap (2026-11-06) — CP200/CP300 positioned against Cortex-A78 and Neoverse N2; “20 %+ per generation” framing
- Analysis (2026-11-06) — RVA23 / server platform spec 1.0 context, SHD Group SoC shipment forecast cited by the vendor
- KUBUDS RISC-V daily digest (2026-11-06) — event timeline (summit 18–20 October) and cross-check of claims