RISC-V has a matrix-extension problem, and it is not a silicon problem — it is a
vocabulary problem. Three competing proposals are currently in flight at RISC-V International:
the Integrated Matrix Extension (IME), the Vector-Matrix Extension (VME)
— the Zvt proposal in the ISA manual — and the Attached Matrix Extension
(AME), all sitting next to the ratified RVV vector extension. Any kernel author
who wants to write tiled matrix code today has to pick a proposal, pick a register model, and accept that
the code may not survive the standardization fight. This week SpacemiT — the company
behind the K1/K3 SoCs — published an attempt to decouple that choice: RVISA (RISC-V Virtual
ISA), a working draft that defines virtual instruction syntax for scalar, vector, and matrix
computation, with versioned mappings onto each of the three proposals.
What Shipped This Week
The repository spacemit-com/rvisa
was created on 4 November 2026 and received its substantive content in a rapid sequence on
5 November: the MkDocs documentation site and Pages release flow, the first examples of the
RVISA notation with the initial IME/VME/AME mappings, and then a pull request defining the programmer-facing
tile layout and blocked matrix multiplication. The specification is versioned as 0.2.0,
licensed Apache-2.0, and self-describes as a working draft —
explicitly not a ratified RISC-V International specification.
Honesty first. Everything below describes a draft, not a
product. The notation cannot be compiled or executed in this revision (source-language
frontends are out of scope), the MLIR dialect is TBD, executable conformance tools are
TBD, and the three registered target profiles have their verification layers recorded
— with not-run where no evidence exists. Read it as a proposal with unusually explicit
contracts, not as usable software.
| Item | Value (as published) |
| Specification version | 0.2.0 (VERSION file; CHANGELOG tracks the draft as “Unreleased”) |
| Status | Working draft — not a ratified RISC-V International specification |
| License | Apache License 2.0 (docs and code) |
| Scope | Virtual instruction syntax for RISC-V scalar, vector, and matrix computation |
| Abstraction slot | Above NVIDIA PTX, below NVIDIA cuTile |
| Notation basis | Modeled on RVV Intrinsics: typed operations, operand roles, register-group constraints |
| Target profiles | ime-main, vme-main, ame-main (draft, registered) |
| Reference workloads | 6 — vector addition, GEMM, epilogues, block scaling, convolution, attention (.rvisa source forms only) |
| MLIR dialect | Design documents present; implementation TBD |
| Executable conformance | TBD; spec-level positive/negative cases defined |
| RFCs | 0001 (instruction/register mapping contracts), 0002 (core operation contracts), 0003 (programmer-expressible tile layouts) — all drafts |
| Repository timeline | Created 2026-11-04; substantive commits 2026-11-05 |
The Slot Between PTX and cuTile
The positioning is stated plainly in the README, and it is the most consequential sentence in the
repository: RVISA targets an abstraction level “slightly above NVIDIA PTX and below NVIDIA
cuTile”. Unpacking that sentence explains the whole design:
- Below a language. RVISA is not a frontend. Python or C++ host forms may accompany a
workload, but source-language frontend implementations are explicitly excluded from this revision. There
is no parser you can call, no compiler you can run.
- Above an encoding. RVISA does not define instruction encodings for any chip. Instead,
“target mappings” are versioned profiles that declare, item by item, which register classes,
register groups, and instruction sequences a given implementation lowers to — each pinned to a
specific upstream revision with its own verification evidence.
- The PTX analogy. NVIDIA’s PTX succeeded because it is a stable virtual
target that survives GPU generational churn: compiled once, re-optimized per generation. RVISA is
betting that RISC-V matrix work needs the same layer — one notation that outlives whichever of IME,
VME, or AME wins.
In notation form it follows the RVV Intrinsics style: typed operations, operand roles, and constraints
on register groups — deliberately familiar to anyone who has written RVV kernel code.
Three Matrix Extensions, One Notation
The reason a virtual layer is even worth proposing is the state of the upstream matrix-extension
landscape. RVISA explicitly roots its design in all three proposals rather than picking one:
Each mapping profile in mappings/ identifies its exact upstream revision,
its register and operand requirements, and its verification evidence. That last part matters: the
CHANGELOG states that each of the first three profiles “pins an upstream commit, one supported
element tuple, a single microtile geometry, the register classes and groups, one lowering rule family,
and the numerical contract, and records every verification layer explicitly (not-run where
no evidence exists)”. This is a deliberately conservative format — a profile cannot silently
narrow its claims.
Tiles, Linear Layouts, Blocked MMA
The core object is the tile — a value spanning 1–N dimensions that the
notation treats as a first-class operand. Three mechanics define how tiles map onto real registers:
- Register priority. Tile mappings prioritize target registers: vector registers and
register groups first, matrix operand registers, then accumulator registers where available. A tile can
span multiple registers or be partitioned to match instruction geometry.
- Linear layouts. Physical placement is derived from linear layouts: explicit
conversions and load/store/MMA/
textract anchors determine how logical element positions
become physical register positions. Reshape and axis permutation must preserve element locations.
- Blocked MMA. The
spec/operations/matrix/blocked-mma.md chapter defines
matrix multiply on tiles that are larger than any single register group, decomposed as blocked
computation — the CHANGELOG records an axis-bound MMA with a 32×32 / 8×8×8
example plus manual fragment expansion as conformance cases.
The primitive set also includes textract/tinsert for tile subvalues and
tselect for boolean elementwise selection, compiler-managed transfers between tiles and
memory, and convolution-window loading — i.e., the pieces a real kernel needs beyond the multiply
itself.
The Expressibility Requirement
The single sharpest design decision is buried in
spec/core/tile-optimization.md: every optimization of an RVISA tile computation must
be expressible by the programmer.
Why this matters. The classic failure mode of matrix-compilation
stacks is silent transformation: the compiler fuses, reorders, or re-tiles your computation in a way you
cannot see, cannot reproduce by hand, and cannot port — so when the next chip arrives, the kernel
is rewritten from scratch. By requiring that any optimization the compiler performs has a
programmer-writable equivalent in the notation, RVISA makes the optimization surface a contract
rather than an implementation artifact. The same wording appears in the CHANGELOG: “Require
programmer-expressible equivalents of tile optimizations; define rank-1 through rank-N linear carriers and
legal register-group decomposition.”
This also sets the boundary for what profiles may do: mapping guarantees define the legal sequences,
grouping, allocation constraints, and transfers for each profile, and they “allow implementation
choices within those constraints” — but a profile cannot claim conformance while quietly
relying on a transformation the programmer could not have written.
What a Target Profile Pins Down
| Profile element | Contract |
| Upstream revision | Pinned commit of the respective extension repo (IME / VME / AME) |
| Element types | One supported element tuple per profile |
| Microtile geometry | A single geometry per profile |
| Registers | Register classes and groups, with tile register priority |
| Lowering | One lowering rule family |
| Numerics | An explicit numerical contract |
| Verification | Every layer recorded, not-run where no evidence exists |
The specification separates three things that are usually conflated: the core
specification (observable behavior), the MLIR dialect (compiler representation
— responsibilities documented, implementation TBD), and target profiles (the
conditions under which the behavior is implementable on particular silicon). The status page is explicit
that “a defined mapping requirement does not establish a supported target or confer the profile
maturity state stable.”
What Is Defined vs. What Is Open
The status page maintains an unusually detailed “open contracts” table. Selected entries:
| Area | Unresolved |
| Vector types | Public fractional-LMUL and mask spelling; aggregate/grouped bridges |
| Configuration | Dynamic-request return value of vconfig; behavior outside the defined full/tail domain |
| Extended stores | Complete address contracts for vsstore and tsstore |
| Tile reductions | treduce_sum axes/result/singletons/numerics; complete broadcast tag notation |
| Matrix numerics | Rounding/special-value details for tmma/vmadot; FMA/format policies for scaled accumulation |
| Low precision | Complete format/conversion coverage; packed FP4 memory ABI |
| Control / ABI / runtime | Negative-step range semantics, cross-compilation-unit calling conventions, host invocation, cross-program synchronization |
| IR / metadata | MLIR dialect encodings, versioned metadata schemas, cost selection |
Two governance rules are worth quoting because they are rare in vendor drafts: “An open contract
must be resolved explicitly in its primary definition and reflected in affected IR descriptions, mappings,
demos, and verification cases”, and “A profile cannot silently choose an unqualified
interpretation and claim general conformance.”
How to Read It: Convergence Layer or Vendor Proposal?
The fair reading has two halves.
The structural argument is sound. RISC-V’s openness produced three competing
matrix-extension proposals; some convergence layer above the hardware is how this resolves without every
kernel being written three times. The same week Quintauris — the Bosch/Infineon/Nordic/NXP/Qualcomm/ST
joint venture — shipped EM-Altair,
a binary-compatibility platform for MCU-class RISC-V. The two moves are the same logic at opposite ends
of the market: MCU-side convergence via a fixed execution environment, AI-matrix-side convergence via a
virtual notation and mapping contracts. Both attack fragmentation by pinning a stable software target.
The asymmetry is equally real. EM-Altair speaks for six major semiconductor vendors;
RVISA is a single-vendor working draft, four days old at the time of writing, with no
compilable notation, no executable conformance, no MLIR implementation, and no implementation evidence
behind its profiles. Whether it becomes infrastructure or remains a position paper depends on one thing:
whether the upstream matrix-extension discussions (IME, the Zvt PR, AME) adopt the mapping
tables — or at least whether other matrix-hardware vendors register their own profiles against the
same contract format. The specification is engineered so that either outcome is cheap to detect: a
profile either fills in its evidence or it does not.
For kernel authors and compiler engineers the pragmatic takeaway: the contract format — pinned
upstream revisions, single element tuple, single microtile geometry, explicit not-run
verification layers — is a useful template today regardless of RVISA’s fate, because
it is the first time the three matrix proposals have been described in one comparable notation. The
six reference workloads (vector add, GEMM, epilogues, block scaling, convolution, attention) with
.rvisa source forms make that comparison concrete.
中文摘要
2026 年 11 月 4–5 日,进迭时空(SpacemiT)在 GitHub 开源了 RVISA(RISC-V 虚拟指令集架构),工作草案 v0.2.0,Apache-2.0 许可。它为 RISC-V 的标量、向量与矩阵计算定义虚拟指令记法,定位明确:在 NVIDIA PTX 之上、cuTile 之下。
要解决的问题:RISC-V 上游同时存在三套矩阵扩展提案——IME(集成矩阵扩展)、VME(ISA 手册 PR #2997 的 Zvt 提案)与 AME(附着矩阵扩展),外加已批准的 RVV。内核作者写一份分块化代码就得选边站。RVISA 用一套记法(参照 RVV Intrinsics 风格:类型化操作、操作数角色、寄存器组约束)覆盖这三套提案,再通过版本化目标映射(target profile)逐条声明落到哪些寄存器类、寄存器组与指令序列。首批注册的三个 profile(ime-main/vme-main/ame-main)各自锁定上游提交 commit、单一元素类型组、单一微分块几何、一族降级规则与数值契约,并逐层记录验证状态——无证据处标注 not-run。
最锐利的设计决策是"可表达性要求":编译器对 tile 计算的任何优化,都必须能被程序员在记法层手写出来。这直接针对矩阵编译栈最常见的失败模式——编译器偷偷向量化、换芯片就得重写——把优化从"实现细节"变成"契约条款"。
诚实前提:这是单家厂商的工作草案,不是 RISC-V International 批准规范;记法本版不能编译执行;MLIR dialect 与可执行一致性工具均标注 TBD;三个 profile 均无实现证据。它能否成为基础设施,取决于上游三条矩阵扩展讨论线是否接受这套映射表。与同期 Quintauris 六大厂商合资公司发布的 EM-Altair MCU 二进制兼容平台互为照应:同样的逻辑——用稳定的软件目标收敛碎片化。
Краткое резюме (RU)
4–5 ноября 2026 года SpacemiT опубликовала на GitHub RVISA (RISC-V Virtual ISA) — рабочий черновик v0.2.0 под Apache-2.0. Это слой виртуальных инструкций для скалярных, векторных и матричных вычислений RISC-V, сознательно размещённый над NVIDIA PTX, но под cuTile.
Мотив: в RISC-V одновременно существуют три конкурирующих предложения матричных расширений — IME, VME (Zvt, PR #2997 в руководстве ISA) и AME, плюс ратифицированный RVV. RVISA описывает все три в одной нотации (стиль RVV Intrinsics) и связывает их с аппаратной реализацией через версионированные target-профили: ime-main, vme-main, ame-main. Каждый профиль фиксирует commit вышестоящего репозитория, один набор типов элементов, одну микротайл-геометрию, классы и группы регистров, одно семейство правил понижения и численный контракт, а слои верификации маркируются not-run при отсутствии доказательств.
Ключевое требование выразимости: любая оптимизация tile-вычислений должна быть выражаема программистом в нотации — оптимизация становится контрактом, а не скрытой работой компилятора.
Честность: это черновик одного вендора, не ратифицированная спецификация RISC-V International; нотация не компилируется, MLIR и исполняемая проверка соответствия — TBD. Это своеобразный параллель EM-Altair от Quintauris: стабильный программный целевой объект против фрагментации.
Resumen (ES)
Del 4 al 5 de noviembre de 2026, SpacemiT publicó en GitHub RVISA (RISC-V Virtual ISA), un borrador de trabajo v0.2.0 bajo Apache-2.0 que define sintaxis de instrucciones virtuales para cálculo escalar, vectorial y matricial en RISC-V, situado deliberadamente por encima de NVIDIA PTX y por debajo de cuTile.
Motivación: RISC-V tiene tres propuestas simultáneas de extensión de matriz — IME, VME (Zvt, PR #2997 del manual ISA) y AME — además del RVV ratificado. RVISA describe las tres en una sola notación (estilo RVV Intrinsics) y las conecta con el hardware mediante perfiles de destino versionados: ime-main, vme-main y ame-main. Cada perfil fija el commit upstream, una sola tupla de tipos de elemento, una sola geometría de microtile, las clases y grupos de registros, una familia de reglas de lowering y un contrato numérico, registrando cada capa de verificación con not-run donde no hay evidencia.
La decisión más aguda es el requisito de expresabilidad: toda optimización de un cálculo de tiles debe poder escribirse por el programador en la notación — la optimización pasa de ser un artefacto del compilador a ser un contrato.
Honestidad: es un borrador de un solo proveedor, no una especificación ratificada por RISC-V International; la notación no se compila ni ejecuta en esta revisión, el dialecto MLIR y las herramientas de conformidad ejecutable son TBD, y los tres perfiles carecen de evidencia de implementación. Es el paralelo de matriz-AI al EM-Altair de Quintauris: un objetivo de software estable contra la fragmentación.
Résumé (FR)
Les 4 et 5 novembre 2026, SpacemiT a publié sur GitHub RVISA (RISC-V Virtual ISA), un avant-projet v0.2.0 sous Apache-2.0 définissant une syntaxe d'instructions virtuelles pour le calcul scalaire, vectoriel et matriciel RISC-V, positionnée délibérément au-dessus de NVIDIA PTX et en dessous de cuTile.
Rationale : RISC-V porte simultanément trois propositions concurrentes d'extension matricielle — IME, VME (Zvt, PR #2997 du manuel ISA) et AME — auxquelles s'ajoute le RVV ratifié. RVISA décrit les trois dans une même notation (style RVV Intrinsics) et les relie au matériel via des profils cibles versionnés : ime-main, vme-main, ame-main. Chaque profil épingle le commit amont, un seul tuple de types d'éléments, une seule géométrie de microtile, les classes et groupes de registres, une famille de règles de lowering et un contrat numérique, chaque couche de vérification étant marquée not-run en l'absence de preuve.
La décision la plus tranchante est l'exigence d'expressibilité : toute optimisation d'un calcul de tiles doit être exprimable par le programmeur dans la notation — l'optimisation devient un contrat et non un effet de bord du compilateur.
Honnêteté : c'est un avant-projet d'un seul vendeur, pas une spécification ratifiée par RISC-V International ; la notation n'est ni compilable ni exécutable dans cette révision, le dialecte MLIR et les outils de conformité exécutable sont TBD, et les trois profils n'ont aucune preuve d'implémentation. C'est le pendant matriciel de l'EM-Altair de Quintauris : une cible logicielle stable contre la fragmentation.
Kurzfassung (DE)
Vom 4. bis 5. November 2026 veröffentlichte SpacemiT auf GitHub RVISA (RISC-V Virtual ISA), einen Arbeitsentwurf v0.2.0 unter Apache-2.0, der virtuelle Instruktionsnotation für skalare, vektorielle und Matrixberechnung auf RISC-V definiert — bewusst über NVIDIA PTX und unter cuTile angesiedelt.
Motivation: RISC-V trägt drei konkurrierende Matrix-Erweiterungsvorschläge gleichzeitig — IME, VME (Zvt, PR #2997 im ISA-Handbuch) und AME — neben dem ratifizierten RVV. RVISA beschreibt alle drei in einer Notation (Stil der RVV-Intrinsics) und verbindet sie über versionierte Zielprofile mit der Hardware: ime-main, vme-main, ame-main. Jedes Profil pinnt den Upstream-Commit, genau eine Elementtyp-Tupel, genau eine Mikrotile-Geometrie, Registerklassen und -gruppen, genau eine Lowering-Regelfamilie und einen numerischen Vertrag; jede Verifikationsebene wird explizit protokolliert — not-run, wo keine Evidenz existiert.
Die schärfste Designentscheidung ist die Ausdrucksbarkeitsanforderung: jede Optimierung einer RVISA-Tile-Berechnung muss vom Programmierer in der Notation ausdrückbar sein — Optimierung wird vom Implementierungsdetail zum Vertrag.
Ehrlichkeit: dies ist ein Arbeitsentwurf eines einzigen Anbieters, keine von RISC-V International ratifizierte Spezifikation; die Notation ist in dieser Revision nicht kompilierbar oder ausführbar, MLIR-Dialekt und ausführbare Konformitätswerkzeuge sind TBD, und die drei Profile haben keine Umsetzungsevidenz. Es ist das Matrix-Pendant zu Quintauris EM-Altair: ein stabiles Softwareziel gegen Fragmentierung.
خلاصه (FA)
در ۴ تا ۵ نوامبر ۲۰۲۶، شرکت SpacemiT RVISA (RISC-V Virtual ISA) را در GitHub منتشر کرد: نسخه کاری v0.2.0 با مجوز Apache-2.0 که نحو دستورالعملهای مجازی را برای محاسبه اسکالر، برداری و ماتریسی در RISC-V تعریف میکند و ارداده است بالاتر از NVIDIA PTX و پایینتر از cuTile.
انگیزه: در RISC-V همزمان سه پیشنهاد رقیب توسیعهی ماتریسی وجود دارد — IME، VME (Zvt، PR #2997 در دستور ISA) و AME — همراه RVV تصویبشده. RVISA همهی را در یک نحو (با سبک RVV Intrinsics) توصیف میکند و با پروفایلهای هدف نسخهبندیشده به سختافزار پیوند میدهد: ime-main، vme-main، ame-main. هر پروفایل commit آپدریم، یک مجموعه نوع ارزش، یک هندسه میکروتایل، کلاسها و گروههای ثبتها، یک خانواده قوانین lowering و یک قرارداد عددی را ثبت میکند، و هر لایه احراز در حذف شواهد با not-run نشانه میشود.
دقیقترین تصمیم طراحی الزامهی بیانپذیری است: هر بهینهسازی روی محاسبهی tile باید توسط برنامهنویس در نحو قابل بیان باشد.
صداقت: این یک پیشنویس تکدارنده است، نه مشخصاتهای تصویبشدهی RISC-V International، نحو در این نسخه کامپایل نمیشود، dialect MLIR و ابزارهای سریاندرسی اجرایی TBD هستند، و سه پروفایل بدون شاهد پیادهسازیاند. این، همراه EM-Altair از Quintauris، هدف نرمافزاری پایدار در برابر پراکردگی است.