The Bottleneck Was Never the ISA: Andes Automates Custom Instructions in COPILOT V7
Every RISC-V vendor sells the same differentiator against Arm: you can add your own instructions. What none of them put on the slide is the bill. A customer instruction is not one artefact — it is at least four: the datapath change in the CPU RTL, the encoding that has to be reserved and documented, the compiler back end that has to emit it, and the simulator model that has to reproduce it so software can be verified before silicon exists. Historically that has meant months of work by people who understand pipelines, requested by an algorithm team that understands none of it.
On 10 November 2026 in Hsinchu, Andes Technology (TWSE: 6533) shipped its latest attempt to delete that bill: the next generation of Andes Automated Custom Extensions (ACE), driven by COPILOT V7. The claim is narrow enough to be tested: a custom instruction is generated into the CPU RTL, the compiler and the simulator from a single description, instead of being hand-built three times by three teams.
What Andes Announced
The release has three parts that matter to a chip team: the framework, the tool, and the reach. The
framework (ACE) is the mechanism by which a customer’s operation becomes part of the core’s
native execution flow rather than a block hanging off the bus. The tool (COPILOT V7) is what turns the
description into the three artefacts. The reach is the part that changes buying decisions: ACE now
spans the full Andes CPU line, from the 3-stage-pipeline D23 MCU core to the out-of-order
application processor AX66 — previously it was not something a small MCU
customer could assume.
| Item | What the 10 November 2026 announcement states |
|---|---|
| Framework | Andes Automated Custom Extensions (ACE), next generation; integrates customer instructions into the CPU RTL, the compiler and the simulator |
| Tool | COPILOT V7; described as significantly reducing the complexity that traditional CPU customisation carries |
| Core coverage | Full CPU family — from D23 (3-stage pipeline, MCU class) to AX66 (out-of-order application class) |
| Stated use cases | Low-latency hardware control via ACE Custom Ports (ACP) to
external blocks such as an NPU; AI/DSP special-function maths (exp, sin,
cos); additional data types (FP8, INT4) |
| Mechanism claim | Custom instructions execute inside the core’s native pipeline — removing the latency, software overhead and memory bottleneck that Andes attributes to memory-mapped coprocessors; a “background processing” mechanism is cited for true parallel acceleration |
| Track record | 20 years of supplying custom-instruction support for AndesCore; >20 billion cumulative Andes-Embedded SoC shipments as of end-2025 (all cores, not ACE specifically) |
COPILOT V7: The Five Additions
Version numbers hide the interesting part, so here is the delta stated in the release. Each item answers a question that a real customer project eventually asks.
| New in V7 | What it does | The question it answers |
|---|---|---|
| ACE_RVV | Adds ACE customisation on top of the high-performance RVV SIMD engine — i.e. custom vector instructions, not just scalar ones | “My kernel is already vectorised; a scalar op will not help me” |
| FRF access | ACE logic can access floating-point registers directly | “Do I have to round-trip my operand through integer registers or memory?” |
| 64-bit ACE instructions | Doubles the number of encoding bits available to a custom instruction | “My op needs more immediate bits / more operand specifiers than a 32-bit slot allows” |
| Heterogeneous ACE configuration | Cores in the same cluster can each implement
only the ACE instructions they need — e.g. some cores get a high-performance exp(x),
others do not | “I want the accelerator instruction where it pays and not pay area for it everywhere” |
| Quad widening and narrowing | ACE_RVV instructions can perform 4× width data-format conversion | “My datapath is wider than the element type the model produced” |
Three Ways an Operation Enters a Core
The release bundles three integration paths that are often conflated in “we support custom instructions” marketing. They are architecturally different and have different verification burdens.
| Path | How it works | Where it fits | What you must verify |
|---|---|---|---|
| ACP custom port | ACE logic talks to an external block (NPU, accelerator) through a dedicated port driven by the core | Tightly coupled control of a neighbour block; low-latency handshakes | Handshake timing, back-pressure, and what the core does while waiting |
| Special-function instruction | A new opcode executes inside the pipeline
— exp, sin, cos, FP8/INT4 conversions | Fixed-latency maths in a hot loop; replacing a library call with one instruction | Numerical conformance against the library you replaced, and exception/flag behaviour |
| ACE_RVV custom vector op | A custom instruction added to the RVV SIMD engine itself | Per-element work where the win comes from vector width, not from a scalar shortcut | Tail handling, masking, vl behaviour, and the vector-state context save path |
Andes draws the contrast with memory-mapped coprocessors sharply, and the argument is sound on its own terms: a coprocessor reached through a store/load pair pays address computation, bus transaction and fence costs on every call, and its operands have to be visible in memory. Putting the operation in the core’s native execution flow removes that tax — which is exactly why the “special-function instruction” path exists at all for something as mundane as a sine.
The Evidence Behind It
This is not a slideware framework with no users. Andes has supplied custom-instruction support for two decades, and the release names three customers with concrete descriptions.
- Meta MTIA — the strongest data point, because it is a hyperscaler’s production accelerator with a published paper behind it. Meta describes cores built on the AX25-V100 as “single-issue, in-order... with a five-stage pipeline... heavily customised for the required functionality”, where the customisation includes dedicated interfaces, dedicated registers, custom instructions and custom exception handling (the quote is reproduced from MTIA: First Generation Silicon Targeting Meta’s Recommendation Systems, ISCA 2023).
- Phison Electronics — states that integrating Andes RISC-V cores with ACE into its architecture produced a flexible, scalable platform for edge and data-centre AI workloads across the performance, power and cost axes.
- Fractile — says it will combine Andes’ high-performance RISC-V vector processors with its in-memory-computing architecture through ACE. Note the tense: will.
The Fine Print: Seven Things to Ask Before You Encode
If the differentiation layer is about to become cheap, the questions move from “can we afford an instruction” to “which instruction, and who can debug it”. These are the ones that decide whether a custom op is an asset or a liability.
| Question | Why it decides the project |
|---|---|
| 1. How does the op appear in the ISA string? | If the custom extension needs a vendor-prefixed string that only the vendor toolchain accepts, you have bought a compiler lock-in along with the instruction. Ask whether upstream binutils/GCC/LLVM can at least disassemble it. |
| 2. Is there a simulator model? | Without an instruction-accurate model the instruction cannot be verified before silicon, which removes most of the point of a late-bound ISA. This is the real deliverable of COPILOT, more than the RTL. |
| 3. What happens on a heterogeneous cluster? | If core 0 has the op and core 1 does not, the same binary is no longer uniformly valid. You need thread affinity, dynamic dispatch, or an OS-visible asymmetry — and your test matrix just doubled. |
| 4. Does the op survive context switches and debug? | A new architectural register means a new save/restore obligation in the kernel and a new register set for the debugger and profiler. 64-bit ACE encodings also mean the disassembler must know the new width. |
| 5. Is ACE_RVV conditional on the vector unit? | Yes by construction —
no RVV engine, no ACE_RVV. On an MCU-class D23 you are in the scalar and ACP paths, which is a
different optimisation problem from a vector kernel. |
| 6. Is the instruction inside the certified envelope? | In automotive terms, a SEooC-style core certification covers a defined configuration. A customer instruction added afterwards sits outside that argument until someone re-argues it — and that cost lands on the chip vendor, not on Andes. |
| 7. Does it outlive the next generation? | A custom instruction is a private ABI with your own software. If the next core cannot express it, the migration is a rewrite. This is the same question that makes proprietary matrix extensions risky in the first place. |
Why It Matters
For a chip team, the practical consequence is that the entry price of differentiation is falling, which moves the scarce resource from HDL engineers to people who can identify which operation is worth an opcode. That is a much better problem to have, and it is the problem Arm customers do not get to have at all.
For a buyer, the test is simple and can be run in a meeting: when a supplier claims a “custom AI instruction”, ask for the compiler flag that emits it, the simulator that models it, and the ISA string it appears in. If all three exist, it is an instruction. If only a datasheet claim exists, it is a coprocessor with better marketing — and you are paying bus latency for it.
For the ecosystem, this is one more data point in the direction we have been tracking all month: value moving off the ISA and onto what sits above it — Onsemi buying Codasip’s low-end line, Quintauris standardising an MCU platform, Infineon putting twelve RISC-V cores into an ASIL D automotive family, MIPS building robot joints on its own cores. Andes is not competing on the base instruction set. It is selling the machine that lets everyone else stop competing on it either.
摘要(中文),2026 年 11 月 10 日,晶心科技(Andes Technology,TWSE:6533)在新竹发布新一代 Andes Automated Custom Extensions(ACE) 开发框架,搭配 COPILOT V7 工具。 核心主张很窄、也很容易被检验:把客户的定制指令从一份描述自动生成到 CPU RTL、编译器与仿真器, 而不是由三个团队各手写一遍。
COPILOT V7 的五项新增:ACE_RVV(在高性能 RVV SIMD 引擎上做定制向量指令,而非只有 标量)、FRF 访问(ACE 逻辑可直接读写浮点寄存器)、64 位 ACE 指令(可用编码位宽 翻倍)、异构 ACE 指令配置(同一集群内各核按需实现不同 ACE 指令)、四倍宽化/窄化 (ACE_RVV 的四倍宽度数据格式转换)。同时 ACE 的覆盖面从三阶段流水线的 MCU 级 D23 一直扩到乱序 应用级 AX66。
结构看点:RISC-V 的优势从来不在基础 ISA(RV64GC 已是几十家共做的商品),而在基础 ISA 之上那一层归你 所有。ACE 押注的是——只有把进入成本从「一个 HDL 项目」降到「一个产品决策」,这一层才会变成真正的市场。所以要看 的是工具链,而不是指令本身。已落地证据:Meta MTIA 基于 AX25-V100 做了专用接口、专用寄存器、定制 指令与定制异常处理(ISCA 2023 论文);群联电子(Phison)与 Fractile(存内计算,未来时)亦在列。
诚实点:全文没有任何基准、主频、面积、功耗数字,COPILOT V7 也没有发布日期与 价格;「不再需要数月修改 RTL 与工具」是厂商表述、无前后对比。Meta 引文出自 2023 年论文,只证明这一级别定制早已 量产,不证明 V7 更快。最需要注意的是「异构 ACE 配置」:同一集群内各核可能不再共享同一套指令集——这是 硬件给软件出的新题(线程迁到不支持该指令的核即非法指令),而稿子里没有一个字的量化。
Кратко (RU). 10 ноября 2026 года в Синьчжу компания Andes Technology (TWSE: 6533) представила новое поколение фреймворка Andes Automated Custom Extensions (ACE) вместе с инструментом COPILOT V7. Заявление узкое и проверяемое: заказная инструкция генерируется из одного описания в RTL процессора, компилятор и симулятор, а не пишется вручную трижды разными командами.
Пять нововведений V7: ACE_RVV (пользовательские векторные инструкции поверх движка RVV, а не только скалярные), доступ к FRF (логика ACE напрямую обращается к регистрам плавающей точки), 64-битные ACE-инструкции (вдвое больше бит кодирования), гетерогенная конфигурация ACE (ядра одного кластера реализуют только нужные им инструкции) и четырёхкратное расширение/сужение для ACE_RVV. Охват расширен на всю линейку — от трёхступенчатого MCU-ядра D23 до внеочередного AX66.
Главная мысль: преимущество RISC-V никогда не было в базовом ISA — RV64GC воспроизводят десятки компаний. Преимущество в том, что слой над базовым ISA юридически ваш. ACE — ставка на то, что этот слой станет рынком только когда вход подешевеет с «проекта на HDL» до «продуктового решения». Смотреть надо на инструменты, а не на инструкцию. Подтверждения: Meta MTIA на ядре AX25-V100 с собственными интерфейсами, регистрами, инструкциями и обработкой исключений (статья ISCA 2023), Phison и Fractile (в будущем времени).
Честная оговорка: в релизе нет ни одного бенчмарка, частоты, площади или энергопотребления, нет даты выхода и цены COPILOT V7. «Больше не нужно месяцами править RTL и инструменты» — утверждение вендора без измерений до/после. Цитата Meta относится к работе 2023 года. Особого внимания заслуживает гетерогенная конфигурация: ядра одного кластера могут не разделять один ISA — поток, попавший не на то ядро, получит недопустимую инструкцию, и количественной оценки этого риска в релизе нет.
Resumen (ES). El 10 de noviembre de 2026, en Hsinchu, Andes Technology (TWSE: 6533) presentó la nueva generación de su marco Andes Automated Custom Extensions (ACE) junto con la herramienta COPILOT V7. La afirmación es estrecha y verificable: una instrucción personalizada se genera desde una única descripción hacia el RTL de la CPU, el compilador y el simulador, en lugar de construirse a mano tres veces por tres equipos distintos.
Las cinco novedades de V7: ACE_RVV (instrucciones vectoriales personalizadas sobre el motor SIMD RVV, no solo escalares), acceso a FRF (la lógica ACE lee y escribe directamente los registros de coma flotante), instrucciones ACE de 64 bits (el doble de bits de codificación), configuración ACE heterogénea (cada núcleo de un clúster implementa solo las instrucciones que necesita) y ensanchamiento/estrechamiento cuádruple. La cobertura abarca ahora toda la familia, desde el núcleo MCU de tres etapas D23 hasta el AX66 fuera de orden.
La lectura estructural: la ventaja de RISC-V nunca fue el ISA base —RV64GC lo implementan una docena de empresas— sino que la capa por encima del ISA base es jurídicamente suya. ACE apuesta a que esa capa solo será un mercado cuando su coste de entrada baje de «un proyecto de HDL» a «una decisión de producto». Hay que mirar las herramientas, no la instrucción. Evidencia: Meta MTIA sobre AX25-V100 con interfaces, registros, instrucciones y manejo de excepciones propios (artículo ISCA 2023), además de Phison y Fractile (en tiempo futuro).
Nota de honestidad: el comunicado no incluye ningún benchmark, frecuencia, área ni consumo, y tampoco fecha de disponibilidad ni precio de COPILOT V7. «Ya no hay que pasar meses modificando el RTL y las herramientas» es una afirmación del proveedor sin medición antes/después. La cita de Meta proviene de un trabajo de 2023. Lo más relevante es la configuración heterogénea: los núcleos de un clúster pueden no compartir un único ISA, y un hilo migrado al núcleo equivocado ejecuta una instrucción ilegal; el comunicado no cuantifica ese riesgo.
Résumé (FR). Le 10 novembre 2026 à Hsinchu, Andes Technology (TWSE : 6533) a présenté la nouvelle génération de son cadre Andes Automated Custom Extensions (ACE) avec l’outil COPILOT V7. L’affirmation est étroite et vérifiable : une instruction personnalisée est générée depuis une description unique vers le RTL du CPU, le compilateur et le simulateur, au lieu d’être construite trois fois à la main par trois équipes.
Cinq ajouts dans V7 : ACE_RVV (des instructions vectorielles personnalisées au-dessus du moteur SIMD RVV, et non plus seulement scalaires), accès FRF (la logique ACE lit et écrit directement les registres flottants), instructions ACE 64 bits (deux fois plus de bits d’encodage), configuration ACE hétérogène (chaque cœur d’un cluster n’implémente que les instructions dont il a besoin) et élargissement/rétrécissement quadruple. La couverture s’étend désormais du cœur MCU à trois étages D23 au AX66 hors-ordre.
Lecture structurelle : l’avantage de RISC-V n’a jamais été l’ISA de base — RV64GC est implémenté par une douzaine d’entreprises — mais le fait que la couche au-dessus de l’ISA vous appartienne juridiquement. ACE parie que cette couche ne deviendra un marché que lorsque son coût d’entrée passera d’« un projet HDL » à « une décision produit ». Regardez l’outillage, pas l’instruction. Preuves : Meta MTIA sur AX25-V100 avec interfaces, registres, instructions et gestion d’exceptions dédiés (article ISCA 2023), plus Phison et Fractile (au futur).
Honnêteté : le communiqué ne contient aucun benchmark, fréquence, surface ni consommation, et aucune date de disponibilité ni prix pour COPILOT V7. « Plus besoin de passer des mois à modifier le RTL et les outils » reste une affirmation du fournisseur sans mesure avant/après. La citation de Meta date de 2023. Le point le plus important est la configuration hétérogène : les cœurs d’un cluster peuvent ne plus partager un ISA unique, et un thread migré sur le mauvais cœur exécute une instruction illégale — risque que le communiqué ne quantifie pas.
Zusammenfassung (DE). Am 10. November 2026 stellte Andes Technology (TWSE: 6533) in Hsinchu die nächste Generation seines Andes Automated Custom Extensions (ACE)-Frameworks zusammen mit dem Werkzeug COPILOT V7 vor. Die Aussage ist eng und prüfbar: Eine kundenspezifische Instruktion wird aus einer einzigen Beschreibung in CPU-RTL, Compiler und Simulator generiert — statt dreimal von drei Teams von Hand gebaut zu werden.
Fünf Neuerungen in V7: ACE_RVV (kundenspezifische Vektorinstruktionen auf der RVV-SIMD-Engine, nicht nur skalare), FRF-Zugriff (ACE-Logik greift direkt auf Gleitkommaregister zu), 64-Bit-ACE-Instruktionen (doppelt so viele Kodierungsbits), heterogene ACE-Konfiguration (jeder Kern eines Clusters implementiert nur die benötigten Instruktionen) und vierfaches Widening/Narrowing. Die Abdeckung reicht nun vom dreistufigen MCU-Kern D23 bis zum Out-of-Order-AX66.
Strukturell: Der Vorteil von RISC-V war nie die Basis-ISA — RV64GC implementieren ein Dutzend Firmen — sondern dass die Schicht über der Basis-ISA rechtlich Ihnen gehört. ACE wettet darauf, dass diese Schicht erst dann ein Markt wird, wenn die Eintrittskosten von „einem HDL-Projekt“ auf „eine Produktentscheidung“ sinken. Entscheidend ist das Werkzeug, nicht die Instruktion. Belege: Meta MTIA auf AX25-V100 mit eigenen Schnittstellen, Registern, Instruktionen und Exception-Behandlung (ISCA-2023-Paper), dazu Phison und Fractile (im Futur).
Ehrlichkeitsprüfung: Die Meldung enthält keinen Benchmark, keine Frequenz, keine Fläche und keine Leistungsaufnahme — und weder Termin noch Preis für COPILOT V7. „Keine Monate mehr für RTL- und Tool-Änderungen“ ist eine Herstelleraussage ohne Vorher/Nachher-Messung. Das Meta-Zitat stammt aus 2023. Am wichtigsten ist die heterogene Konfiguration: Kerne eines Clusters teilen möglicherweise keine einheitliche ISA mehr, und ein Thread auf dem falschen Kern führt eine illegale Instruktion aus — ohne jede Quantifizierung in der Meldung.
خلاصه (FA). در ۱۰ نوامبر ۲۰۲۶، شرکت Andes Technology (TWSE: 6533) در سینژو نسل تازهٔ چارچوب Andes Automated Custom Extensions (ACE) را همراه با ابزار COPILOT V7 معرفی کرد. ادعا محدود و قابل آزمون است: دستور سفارشی از یک توصیف واحد به RTL پردازنده، کامپایلر و شبیهساز تولید میشود، نه اینکه سه بار توسط سه تیم دستی ساخته شود.
پنج افزودهٔ V7: ACE_RVV (دستورهای برداری سفارشی روی موتور SIMD برداری، نه فقط اسکالر)، دسترسی به FRF (منطق ACE مستقیماً به ثباتهای ممیز شناور دسترسی دارد)، دستورهای ۶۴ بیتی ACE (دو برابر بیتهای کدگذاری)، پیکربندی ناهمگن ACE (هر هستهٔ یک خوشه فقط دستورهای مورد نیاز خود را پیاده میکند) و پهنسازی/باریکسازی چهاربرابر. پوشش از هستهٔ سهمرحلهای D23 تا AX66 خارج از ترتیب گسترش یافته است.
نکتهٔ ساختاری: مزیت RISC-V هرگز ISA پایه نبود — RV64GC را دهها شرکت پیاده میکنند — بلکه این است که لایهٔ بالای ISA پایه از نظر حقوقی متعلق به شماست. ACE شرط میبندد که این لایه تنها وقتی بازار میشود که هزینهٔ ورود از «یک پروژهٔ HDL» به «یک تصمیم محصول» کاهش یابد. باید به ابزار نگاه کرد، نه به دستور. شواهد: Meta MTIA بر پایهٔ AX25-V100 با رابطها، ثباتها، دستورها و مدیریت استثنای اختصاصی (مقالهٔ ISCA 2023)، و همچنین Phison و Fractile (به زمان آینده).
هشدار صادقانه: در بیانیه هیچ بنچمارک، فرکانس، مساحت یا توانی نیامده و برای COPILOT V7 نه تاریخ عرضه و نه قیمت اعلام نشده است. «دیگر ماهها صرف تغییر RTL و ابزار نمیشود» ادعای سازنده است بدون سنجش پیش/پس. نقلقول Meta از مقالهٔ ۲۰۲۳ است. مهمترین نکته پیکربندی ناهمگن است: هستههای یک خوشه ممکن است دیگر یک ISA مشترک نداشته باشند و رشتهای که به هستهٔ نادرست مهاجرت کند دستور غیرمجاز اجرا میکند — بیآنکه بیانیه این ریسک را کمّی کند.
Sources
- Andes Technology, press release “Andes晶心科技推出新一代 Andes Automated Custom Extensions”, Hsinchu, 10 November 2026 (official release; primary source for every figure quoted here).
- Meta, “MTIA: First Generation Silicon Targeting Meta’s Recommendation Systems”, ISCA 2023 — quoted in the Andes release; the AX25-V100 customisation description is reproduced from that paper.
- Andes Technology product family, andestech.com — D23 (3-stage MCU) and AX66 (out-of-order application) core positioning.
- Related Open RISC-V coverage: SpacemiT RVISA and the matrix-extension standardisation gap (6 November 2026), Infineon AURIX RC1 (5 November 2026), MIPS Actus (10 November 2026).