Ask anyone who has taped out a multi-core processor what actually broke first, and the answer is rarely the core. It is the thing between the cores. Single-core RISC-V performance has been on a public, checkable curve for three years — per-gigahertz scores, pipeline depths, issue widths, all quotable. What has not been on a public curve is the part that decides whether eight of those cores behave like eight cores or like one core and seven spectators: the cache-coherent interconnect. On 12 November 2026, Alibaba DAMO’s RISC-V brand XuanTie put numbers on that layer for the first time in a while, disclosing the XL-300 high-throughput coherent interconnect.
The published claims are short enough to quote in full: support for up to eight cores on asynchronous links, a 256-bit data path, up to 32 MB of configurable L3 cache, eight-core concurrency linearity above 90%, a bandwidth ceiling roughly double that of the previous-generation XL-200, and a new Direct Memory Transfer path that is said to bypass the conventional route and relieve the throughput bottleneck created by a large L3 combined with high bus latency [1].
That is six claims and zero conditions. This article does three things with them: places them against what XuanTie said about the same IP sixteen months earlier, works out what a “90% linearity” figure can and cannot mean, and turns the whole thing into the checklist a board or server vendor should actually run before designing XL-300 into a product.
| Attribute | Disclosed 12 Nov 2026 | Status |
|---|---|---|
| Cores per interconnect | Up to 8, asynchronous connection | Stated; “asynchronous” not defined (separate clock domains? mesochronous?) |
| Data path width | 256 bit | Stated; no per-direction or per-link split |
| L3 cache | Up to 32 MB, configurable | Stated; no associativity, latency, or partitioning granularity |
| Scaling | “8-core concurrency linearity above 90%” | Stated; no workload, no baseline, no frequency or power condition |
| Bandwidth | ~2× the XL-200 ceiling | Relative only; XL-200 absolute figure never published |
| New feature | Direct Memory Transfer | Named; no programming model, no coherence semantics |
| Process / area / power | — | Not disclosed for XL-300 |
| Licensing and price | — | Not disclosed; XuanTie IP is licensed, terms private |
| Silicon proof | — | No product named as shipping XL-300 at these settings |
None of the gaps are unusual for IP disclosure — coherent interconnect is licensed under NDA to a handful of SoC teams, not sold on a web store. But the gaps determine what the announcement can be used for. It is enough to conclude that XuanTie is still investing in the multi-cluster layer that sits under C930 and C950. It is not enough to conclude that a given customer’s eight-core design will see 90% of anything.
XL-300 is not new. It was already named in July 2025, in a XuanTie briefing on the C930 server core, as the strongest member of the in-house XT-Link interconnect family and the one paired with C930 [2]. The XuanTie product site still lists XL-300 as the multi-cluster interconnect for C930, alongside RVA23, RVV 1.0 with VLEN 256, the H hypervisor extension and CoVE confidential computing [3]. At that briefing the specifications were given as:
| Metric | XuanTie, July 2025 (C930 briefing) | XuanTie, November 2026 |
|---|---|---|
| Cores per cluster | Up to 8 (multiple clusters for more) | Up to 8, asynchronous links |
| L3 capacity | Up to 23 MB | Up to 32 MB |
| vs XL-200, frequency | +20% | Not restated |
| vs XL-200, bandwidth | 2× | “Roughly double” |
| vs XL-200, area | +5% | Not restated |
| QoS | Capacity and bandwidth allocation, per-ID acceleration | Not restated |
| Data path | Not stated | 256 bit |
| Linearity | Not stated | ≥90% at 8 cores |
The two disclosures also differ in kind, not just in value. July 2025 gave relative numbers against XL-200 — frequency, bandwidth, area. November 2026 gives absolute configuration ceilings plus one efficiency number. Neither gives absolute bandwidth in GB/s, which is the number a memory-system architect actually needs.
The phrasing is “8核并发线性度达到90%以上” — eight-core concurrency linearity above 90%. Read literally as marginal efficiency, that means the eighth core delivers at least 90% of what the first core delivered, which implies a speedup of 1 + 7 × 0.9 = 7.3× on whatever workload produced the number, i.e. about 91% parallel efficiency. That is a strong result for any shared-memory machine, and an entirely believable one for a well-designed 8-core cluster running a streaming or embarrassingly parallel kernel.
It is also the kind of number that means very little without its conditions, because linearity is a property of a pair (workload, machine), not of an interconnect. The same fabric can show 95% on STREAM triad and 45% on a pointer-chasing graph workload, with nothing dishonest in either measurement. Three conditions would make the claim checkable, and none were published:
One more structural point: an eight-core ceiling is not an accident of this generation. It is where a single cluster stops being cheap to build, because beyond roughly eight requesters the coherence traffic between them grows faster than the useful work. Going wider means multiple clusters and a second level of coherence — which is exactly what XuanTie says XL-300 supports, and exactly where the unquantified part lives.
The most technically load-bearing sentence in the disclosure is the one about Direct Memory Transfer: a path that bypasses the conventional route to break the throughput bottleneck caused by a large L3 together with high bus latency. That is a real problem statement. Once L3 grows to tens of megabytes, the L3 stops looking like a cache and starts looking like a NUMA-ish memory tier: the hit rate goes up, but so does the latency of a miss path that has to traverse more of the fabric, and the interconnect’s own queueing becomes the limit rather than DRAM bandwidth.
What is not said is how the bypass works and who can use it. Three questions decide whether this is an architectural feature or a marketing name:
Put the XL-300 disclosure next to the rest of the market and a pattern appears. Cores are commoditising faster than fabrics are. XuanTie has C930 at roughly 15 SPECint2006 points per gigahertz and C950 above 22 with a tensor engine; UltraRISC quotes over 22/GHz for UR-CP400; SiFive, Tenstorrent, Ventana and the open-source XiangShan line all publish per-gigahertz numbers that a buyer can line up in a table. Every one of those teams can buy or build a decent core. Far fewer can ship a coherent fabric that holds together at 16, 32 or 64 sockets-worth of requesters, with an IOMMU in the path, without deadlocking under real traffic.
That is why the interconnect is the moat, and why the announcement matters more than its numbers suggest. For anyone building a RISC-V server or a high-core-count edge box, the practical question is not “which core” but “which fabric, licensed on what terms, with what software model”. The core determines your single-thread score. The fabric determines whether you can sell a 32-core machine at all.
| # | Question | Why it changes the answer |
|---|---|---|
| 1 | Linearity on which benchmark, at what problem size? | Bandwidth-bound and coherence-bound workloads diverge by 2× or more |
| 2 | Is 90% marginal (core 8 vs core 1) or aggregate (vs perfect 8×)? | Different denominators; only the second is parallel efficiency |
| 3 | Measured at fixed frequency, or with DVFS engaged? | All-core clock drop can fake linearity |
| 4 | Absolute bandwidth in GB/s, read and write, per link? | “2× XL-200” is unusable without an absolute base |
| 5 | L3 hit latency in cycles, and miss latency to DRAM? | A 32 MB L3 with a slow miss path can be worse than 8 MB |
| 6 | Snoop filter or directory? Capacity in lines? | Determines whether coherence traffic scales or broadcasts |
| 7 | 23 MB or 32 MB in the configuration I am licensing? | Two XuanTie figures 16 months apart do not reconcile |
| 8 | What exactly does Direct Memory Transfer bypass, and can software issue it? | Decides whether it is an architectural feature or a name |
| 9 | IOMMU in the same fabric? Which RISC-V IOMMU spec revision? | PCIe passthrough and confidential VMs depend on it |
| 10 | Area and power for the fabric at 8 cores, on which node? | The fabric, not the core, is where server SoC budgets blow up |
| 11 | Which shipping product runs this configuration today? | IP disclosure is not silicon; ask for a part number |
XuanTie’s XL-300 disclosure is a useful signal wrapped in an unverifiable number. The signal is that the company keeps investing one layer below the core, at the point where RISC-V server parts have historically come apart, and that it is willing to put an efficiency figure — 90% at eight cores, implying 7.3× — into public. The unverifiable part is everything that would let an engineer check it: workload, baseline, frequency, absolute bandwidth, L3 latency, snoop capacity, area, power, price.
For the reader building hardware, the correct response is neither to dismiss the number nor to design around it. It is to convert it into the eleven questions above and send them to the IP vendor. Any answer that comes back with a named benchmark, an absolute bandwidth figure and a shipping part number is worth more than every press release in this segment combined.
2026 年 11 月 12 日,阿里达摩院旗下 RISC-V 品牌玄铁(XuanTie)公开了高吞吐缓存一致性互联 IP XL-300 的规格:最多 8 核异步连接、 256 bit 数据位宽、最大 32 MB 可配 L3、 8 核并发线性度 90% 以上、带宽上限较前代 XL-200 翻倍,并新增 Direct Memory Transfer(直接内存传输) 通路,绕开传统路径,缓解“大 L3 + 高总线延迟”带来的吞吐瓶颈。
一个关键事实:XL-300 不是新东西。 2025 年 7 月的 C930 技术说明会上,它已被称为自研 XT-Link 互联家族中最强的一款,与 C930 搭配, 当时给出的口径是:单 Cluster 最多 8 核、 L3 最大 23 MB、相比 XL-200 频率提升 20%、带宽翻倍、 面积仅增 5%。因此同一份 IP 在 16 个月里出现了 23 MB 与 32 MB 两个不能互相解释的 L3 口径。
“90% 线性度”怎么读:若按边际效率理解,第 8 颗核仍能提供第 1 颗核 90% 的效益,则加速比 = 1 + 7×0.9 = 7.3×,即并行效率约 91%。这是很强的结果,但“线性度”属于(负载,机器)二元组, 不属于互联本身:同一个 fabric 在 STREAM 上可以 95%,在指针跳转 的图计算上可以 45%,两个数字都不算作假。公开材料没给出: 哪个负载、哪个基准(单核还是理想 8×)、什么频率与功耗条件。
最值得关注的一句是 Direct Memory Transfer。 它直面了真实的问题:L3 增大到几十 MB 后,它不再像 cache, 更像一层结构性内存:命中率上升,但 miss 路径要穿越更多 fabric, 互联的队列而非 DRAM 带宽成为瓶颈。但未说明:软件是否可见; 其一致性语义如何(绕开一致性点就必须回答共享行怎么办); 只对加速器生效还是也能提升 CPU 核线性度。
为何重要:单核性能已经可以公开横比(C930 约 15/GHz、C950 >22/GHz、 UR-CP400 >22/GHz),但能做出并交付 16/32/64 个请求者级别的一致性 fabric、 并且 IOMMU 在路径里不死锁的团队很少。考虑到本站在售的 SpacemiT K1 正好停在 8 核这个单 Cluster 的边界上,再往上就必须有第二层 一致性;这一层能否授权、面积与功耗多少,比下一代核更能决定 RISC-V 服务器的形态。
该问什么:文中给出 11 条清单,其中最硬的四条是:线性度基于哪个 benchmark、90% 是边际还是总体、绝对带宽(GB/s)是多少、以及当前哪个 量产产品跑的是这个配置。IP 公开不等于硅片,请要产品型号。
12 ноября 2026 бренд XuanTie (структурное подразделение Alibaba DAMO) раскрыл характеристики когерентного интерконнекта XL-300: до 8 ядер на асинхронных линках, шина 256 бит, конфигурируемый L3 до 32 МБ, линейность на 8 ядрах выше 90%, пропускная способность почти вдвое выше XL-200, и новый путь Direct Memory Transfer, обходящий обычный маршрут.
Это не новый IP. В июле 2025 года на брифинге по ядру C930 XL-300 уже назывался сильнейшим в семействе XT-Link, с такими цифрами: до 8 ядер на кластер, L3 до 23 МБ, +20% частоты и 2× пропускной способности против XL-200 при +5% площади. Таким образом, для одного и того же IP существуют две несовместимые цифры L3.
Что значит 90%. Если читать как предельную эффективность, восьмое ядро даёт 90% от первого, ускорение = 1 + 7×0,9 = 7,3×, то есть ~91% параллельной эффективности. Но линейность относится к паре (нагрузка, машина), а не к интерконнекту: одна та же фабрика даёт 95% на STREAM и 45% на разреженном графе. Не указаны ни тест, ни база, ни частота/мощность.
Главное в анонсе — Direct Memory Transfer: обход нужен, когда L3 в десятки МБ перестаёт быть кэшем и становится уровнем памяти. Но неясно: виден ли он софту, каковы его семантики когерентности.
Почему это важно: ядра быстро унифицируются (C930 около 15/GHz, C950 >22/GHz, UR-CP400 >22/GHz), а когерентные фабрики на 16–64 участников с IOMMU строят единицы. Платы на SpacemiT K1, которые продаёт этот сайт, стоят ровно на границе восьми ядер.
11 вопросов вендору приведены в таблице: тест, база, частота, абсолютная пропускная способность, задержка L3, snoop-фильтр, номер партии. IP в анонсе ≠ кристалл.
El 12 de noviembre de 2026, la marca RISC-V XuanTie (de Alibaba DAMO) publicó las cifras del interconectado coherente XL-300: hasta 8 núcleos en enlaces asíncronos, bus de 256 bits, L3 configurable de hasta 32 MB, linealidad superior al 90% con ocho núcleos, techo de ancho de banda casi el doble del XL-200, y un nuevo camino Direct Memory Transfer que evita la ruta convencional.
No es IP nueva. En julio de 2025, durante la presentación del núcleo C930, XuanTie ya lo describía como el miembro más potente de la familia XT-Link, con hasta 8 núcleos por clúster, L3 de hasta 23 MB, +20% de frecuencia y 2× de ancho de banda frente al XL-200 con solo +5% de área. Dieciséis meses después, el mismo IP aparece con 32 MB: dos cifras que no se reconcilian sin una hoja de datos que XuanTie no ha publicado.
Qué significa el 90%. Leído como eficiencia marginal, el octavo núcleo aporta el 90% del primero, lo que implica una aceleración de 1 + 7×0,9 = 7,3×, es decir, cerca del 91% de eficiencia paralela. Es un resultado fuerte, pero la linealidad pertenece al par (carga, máquina): el mismo fabric puede dar 95% en STREAM y 45% en un grafo con saltos de puntero. No se publicó ni la carga, ni la base, ni las condiciones de frecuencia y consumo.
Lo más interesante es Direct Memory Transfer, que ataca un problema real: con decenas de MB, la L3 deja de comportarse como caché y pasa a ser un nivel de memoria. Falta saber si el software puede usarlo y cuáles son sus semánticas de coherencia.
Por qué importa: los núcleos se commoditizan (C930 ~15/GHz, C950 >22/GHz, UR-CP400 >22/GHz) mientras los fabricantes capaces de entregar un fabric coherente para 16, 32 o 64 solicitantes con IOMMU en el camino siguen siendo pocos. Las placas SpacemiT K1 que vende este sitio están justo en el límite de ocho núcleos. El artículo cierra con once preguntas que conviene enviar al proveedor de IP antes de fiarse de una cifra de linealidad.
Le 12 novembre 2026, la marque RISC-V XuanTie (Alibaba DAMO) a publié les caractéristiques de l’interconnexion cohérente XL-300 : jusqu’à 8 cœurs sur liens asynchrones, un bus de 256 bits, un L3 configurable jusqu’à 32 Mo, une linéarité supérieure à 90 % à huit cœurs, un plafond de bande passante presque doublé par rapport au XL-200, et un nouveau chemin Direct Memory Transfer contournant la route classique.
Ce n’est pas une IP nouvelle. En juillet 2025, lors de la présentation du cœur C930, XL-300 était déjà décrit comme le plus puissant de la famille XT-Link : jusqu’à 8 cœurs par cluster, L3 jusqu’à 23 Mo, +20 % de fréquence et 2× la bande passante face au XL-200 pour seulement +5 % de surface. Seize mois plus tard, la même IP est annoncée à 32 Mo. Les deux chiffres ne se réconcilient pas, et aucune fiche technique n’a été publiée.
Que vaut le « 90 % » ? Lu comme une efficacité marginale, le huitième cœur apporte 90 % du premier, soit une accélération de 1 + 7×0,9 = 7,3×, c’est-à-dire environ 91 % d’efficacité parallèle. Mais la linéarité appartient au couple (charge, machine) : le même fabric peut afficher 95 % sur STREAM et 45 % sur un parcours de pointeurs. Ni charge, ni base, ni conditions de fréquence et de consommation ne sont publiées.
Le point le plus intéressant est le Direct Memory Transfer, qui attaque un vrai problème : à plusieurs dizaines de Mo, le L3 cesse d’être un cache et devient un niveau de mémoire. Reste à savoir si le logiciel peut l’utiliser et quelles sont ses sémantiques de cohérence.
Pourquoi c’est important : les cœurs se banalisent (C930 ~15/GHz, C950 >22/GHz, UR-CP400 >22/GHz) alors que les équipes capables de livrer un fabric cohérent à 16, 32 ou 64 demandeurs avec un IOMMU dans le chemin restent rares. Les cartes SpacemiT K1 vendues sur ce site se situent exactement à la limite des huit cœurs. L’article se termine par onze questions à poser au fournisseur d’IP.
Am 12. November 2026 veröffentlichte die RISC-V-Marke XuanTie (Alibaba DAMO) die Daten des kohärenten Interconnects XL-300: bis zu 8 Kerne an asynchronen Links, 256 Bit Datenbreite, bis zu 32 MB konfigurierbarer L3, Linearität über 90 % bei acht Kernen, fast doppelte Bandbreite gegenüber XL-200 und ein neuer Direct Memory Transfer-Pfad, der die klassische Route umgeht.
Es ist keine neue IP. Im Juli 2025, beim C930-Briefing, wurde XL-300 bereits als stärkstes Mitglied der XT-Link-Familie beschrieben: bis zu 8 Kerne pro Cluster, L3 bis 23 MB, +20 % Frequenz und 2× Bandbreite gegenüber XL-200 bei nur +5 % Fläche. Sechzehn Monate später nennt dieselbe IP 32 MB. Beide Zahlen sind von XuanTie, und ohne Datenblatt ist keine Auflösung möglich.
Was heißt 90 %? Als Grenzeffizienz gelesen liefert der achte Kern 90 % des ersten, also 1 + 7×0,9 = 7,3× Speedup, rund 91 % parallele Effizienz. Das wäre stark — aber Linearität gehört zum Paar (Last, Maschine): dasselbe Fabric zeigt 95 % bei STREAM und 45 % bei Pointer-Chasing. Last, Bezugsbasis, Frequenz- und Leistungsbedingungen wurden nicht genannt.
Die interessanteste Zeile ist Direct Memory Transfer. Das Problem ist real: Ab zweistelligen MB hört L3 auf, Cache zu sein, und wird zu einer Speicherebene. Offen bleibt, ob Software den Pfad adressieren kann und welche Kohärenzsemantik er hat.
Warum das zählt: Kerne werden schneller zur Ware (C930 ~15/GHz, C950 >22/GHz, UR-CP400 >22/GHz), während Teams, die ein kohärentes Fabric für 16, 32 oder 64 Teilnehmer mit IOMMU im Pfad liefern können, selten bleiben. Die hier verkauften SpacemiT-K1-Boards liegen genau an der Acht-Kern-Grenze. Der Artikel endet mit elf Fragen an den IP-Anbieter.
در ۱۲ نوامبر ۲۰۲۶، برند RISC-V پنجرهای XuanTie (وابسته به Alibaba DAMO) مشخصات میانپیوند همدوسی XL-300 را منتشر کرد: تا ۸ هسته روی پیوندهای ناهمگام، گذرگاه ۲۵۶ بیتی، حافظه L3 قابلتنظیم تا ۳۲ مگابایت، خطیبودن بالای ۹۰٪ در هشت هسته، سقف پهنای باند تقریباً دو برابر XL-200 و مسیر تازهٔ Direct Memory Transfer که مسیر متعارف را دور میزند.
این IP تازه نیست. در ژوئیهٔ ۲۰۲۵ و در نشست معرفی هستهٔ C930، همین XL-300 قویترین عضو خانوادهٔ XT-Link توصیف شد: تا ۸ هسته در هر خوشه، L3 تا ۲۳ مگابایت، ۲۰٪ فرکانس بیشتر و دو برابر پهنای باند نسبت به XL-200 با تنها ۵٪ سطح بیشتر. شانزده ماه بعد همان IP با ۳۲ مگابایت اعلام میشود؛ دو عدد از یک شرکت که بدون برگهٔ داده قابل آشتی نیستند.
۹۰٪ یعنی چه. اگر بازده حاشیهای خوانده شود، هستهٔ هشتم ۹۰٪ هستهٔ اول را میدهد، یعنی سرعتگیری ۱ + ۷×۰٫۹ = ۷٫۳ برابر و بازده موازی حدود ۹۱٪. نتیجهٔ قویای است، اما خطیبودن به زوج (بار کاری، ماشین) تعلق دارد: همان بستر روی STREAM میتواند ۹۵٪ و روی پیمایش اشارهگر ۴۵٪ بدهد. نه بار کاری، نه مبنا، نه شرط فرکانس و توان منتشر نشدهاند.
مهمترین جملهٔ اعلامیه Direct Memory Transfer است: وقتی L3 به دهها مگابایت میرسد، دیگر کش نیست و به یک سطح حافظه تبدیل میشود. اما روشن نیست که نرمافزار آن را میبیند یا نه و معنای همدوسیاش چیست.
چرا مهم است: هستهها بهسرعت کالایی میشوند (C930 حدود ۱۵ و C950 و UR-CP400 بالای ۲۲ امتیاز بر گیگاهرتز)، اما تیمهایی که میتوانند بستر همدوس برای ۱۶ تا ۶۴ درخواستکننده با IOMMU در مسیر تحویل دهند اندکاند. بردهای SpacemiT K1 که این سایت میفروشد دقیقاً روی مرز هشت هسته ایستادهاند. مقاله با یازده پرسش از فروشندهٔ IP پایان مییابد.