Skip to content
AevornixAEVORNIX
AI runtimeResearch

Will this model run on RISC-V?

RVStack AI helps developers check, optimise, benchmark and package AI models for RISC-V software environments.

Architecture

Eight stages, upload to deployable bundle.

The pipeline answers one question in stages and stops at the first stage that says no. An unimplemented operator is a cheaper answer than a failed build.

Memory and latency are estimates until the benchmark stage runs on a real target. The page labels them that way, and a benchmark figure will carry the environment it was produced on.

AI runtime pipeline
AI runtime pipeline, 8 stages first to last: 01 Model upload, 02 Format check, 03 Operator scan, 04 Quantisation, 05 Runtime selection, 06 RISC-V test, 07 Benchmark, 08 Deployment export.
  1. Model upload

    A model file enters the pipeline with its intended target profile.

  2. Format check

    Identify the format, its version and the toolchain that exported it.

  3. Operator scan

    Enumerate operators and flag those with no RISC-V implementation.

  4. Quantisation

    Report stored precision and whether a smaller one is viable.

  5. Runtime selection

    Choose a runtime for the format and operator set, or report that none fits.

  6. RISC-V test

    Build runtime and model for the target and confirm they load and execute.

  7. Benchmark

    Record timings and memory from a real run, with the environment named.

  8. Deployment export

    Emit model, runtime and configuration as one installable bundle.

Modules

What each stage is made of.

Ten modules. Seven are at research stage and three are scheduled behind them.

Model intake

Accept a model file and identify its format, version and export toolchain.

Research

Operator scan

List every operator the graph uses and flag those with no RISC-V implementation.

Research

Quantisation check

Report the precision the model is stored at and whether a smaller one is viable.

Research

Runtime selection

Recommend a runtime for the format and operator set, or state that none is suitable.

Research

Memory estimate

Estimate weights, activations and peak working set against a profile budget.

Research

Latency estimate

Estimate inference cost against a profile, stated as an estimate rather than a measurement.

Research

RISC-V build test

Build the runtime and the model for the target and confirm it loads and executes.

Research

Benchmark report

Record real timings from an emulated or physical run, with the environment named.

Planned

Deployment bundle

Package model, runtime and configuration as one installable artefact.

Planned

Store listing

Publish the bundle with its scan results, licence and tested date attached.

Planned

Model targets

Formats and categories under assessment.

Four formats and five model categories. The level column will change per row as the work progresses.

LevelsSupportedRuns, with a published result behind it.TestingRunning in the pipeline; results not yet stable.ResearchBeing assessed. Nothing runs yet.UnsupportedAssessed and ruled out for now.
Model formats and categories, their current assessment level, and what is being assessed.
Model targetTypeLevelWhat is being assessed
ONNXformatResearchBroadest operator coverage to assess; the first format under investigation.
TensorFlow LiteformatResearchMobile-oriented and already quantisation-aware, which suits a constrained profile.
GGUFformatResearchQuantised language models; the runtime question is vector-extension support.
PyTorch exportformatResearchRequires an export step before any RISC-V question can be asked.
Speech modelscategoryResearchRecognition and synthesis at sizes a mobile power budget allows.
Vision modelscategoryResearchClassification and detection; sensitive to available memory bandwidth.
Text modelscategoryResearchSmall language models, dominated by quantisation and memory.
Translation modelscategoryResearchSequence models with a wide operator surface to check.
Recommendation modelscategoryResearchEmbedding-heavy; storage and lookup cost matter more than compute.

Not every model runs on every architecture. The scanner identifies which will and which will not, before development time is committed.