← Back to Blog
RISC-V AI

RISC-V Matrix Extensions in 2026: IME, VME and AME, Plus the VME Ratification Dates

RISC-V AI

Why matrix, and why it is late

RISC-V ratified a vector extension (RVV) in 2021. Vectors handle element-wise work well, but the dominant operation in neural inference is the outer product of two tiles, and doing that with vector instructions costs register pressure and accumulator traffic. A matrix extension folds the outer product into one instruction with a dedicated accumulator.

RISC-V deliberately did not standardise one matrix extension. At Summit Europe 2025 in Paris the effort was restructured into a family.

The three tracks, plus one

The split exists because RISC-V spans 32-bit microcontrollers to datacenter parts. One extension sized for the top of that range does not fit the bottom, and forcing a single answer would produce the fragmentation the profiles work was meant to end.

VME: the one with a schedule

VME is the furthest along. Its published ratification plan, as relayed by the Shanghai Jiao Tong University ISC RISC-V group's technical digest, sets these dates:

MilestoneDate
Specification freeze2026-09-25
Public review opens2026-10-09
TSC approval2026-11-04
Board of Directors approval2026-11-18

The Unprivileged ISA Committee ran a ratification-plan approval vote through OpaVote in July 2026, closing 2026-07-31.

The action item: freeze is 18 days out from publication of this draft. If your workloads depend on matrix instruction semantics, the window to get comments in is the public review starting 2026-10-09, not after.

IME and VME are converging, not competing

A Summit Europe 2026 talk by Philipp Tomsich (VRULL; chair of the RISC-V Applications & Tools Committee, RISC-V board member, vice-chair of the TSC, and principal editor of IME) reports that IME and VME are converging on specification freeze. The architectural decisions named: algebraic tile geometry that scales with VLEN, so one instruction shape works across different vector widths; deliberate reuse of RVV state, so software needs no parallel register-file abstraction; and dedicated accumulator registers where implementations need higher computational intensity.

The talk also states that work is starting to unify IME and VME through a common LLVM/MLIR lowering path, giving compilers and AI/ML frameworks one abstraction targeting both. That is the most consequential item here: two matrix extensions with one compiler path do not fragment the software stack. Two with separate paths do.

AME is the one still under review pressure

In July 2026, Krste Asanović raised two objections to the AME ratification plan: Spike support was not listed in the simulator section, and no hardware proof-of-concept was listed. On the second, his position is that a credible entity should indicate it has built something similar — even for a different ISA — to show the extension is efficiently implementable; the implementation itself can stay private. Thomas Gall added that Spike support also matters for certification, which wants test suites checked against as many simulators as possible. The plan was updated 2026-07-28 to add both.

Read that as a process signal, not a blocker: an extension adding a large independent register file is a bigger silicon commitment than one reusing vector state, and the committee wants evidence before the spec becomes immutable.

What this means if you are picking silicon today

Nothing here ships yet. No ratified matrix extension means no portable matrix intrinsics, so any matrix-accelerated RISC-V part you buy today uses a vendor extension with vendor toolchain support — defensible for a fixed-function product, poor if you need to move the codebase. The dates to plan against are the VME freeze (2026-09-25), public review (2026-10-09), and BoD approval (2026-11-18).

What is not disclosed

Sources

Verification notes