Chipsterfounders@chipster.ai

Formal verification that learns from every run

Quantitative Based Formal Verification for Chip design at the RTL level

Chipster reads a spec and RTL, writes SystemVerilog assertions, and checks them with a formal tool. Assertions that fail stay in the record. Simulation explains why, and that lesson is available on the next design. The engineer reviews what ships.

How it works

One loop, from this design to the next.

Chipster does not ask a model for a single assertion and stop. It searches, proves what it can, keeps the counterexamples, and uses them the next time.

  1. 01

    Read

    Feed Chipster with Spec and the RTL with extra context for grounding.

  2. 02

    Extraction

    Chipster will generate requirements for every signal using a Knowledge Graph. Accept what requirements you want assertions for.

  3. 03

    Assertion Generation

    Chipster will start its MCTS algorithm for assertion generation using your EDA formal backend. It will keep the best assertions.

  4. 04

    Continual Learning

    Dropped assertions are replayed in simulation to extract useful insights. These insights make future assertion generation better, prevent repeated mistakes, and surface insights to the engineer.

  5. 05

    Remember

    Start each new design with knowledge from previous runs, enabling faster sign-off and higher coverage.

Example Design

An out-of-order CPU. The property: an instruction must not commit before every older instruction has committed.

Example design

Out-of-order CPU

Complex, concurrent

Target property

An instruction must not commit before all older instructions have committed.

1 Generate

assert property (
  commit_valid |->
  rob_head_idx ==
    commit_idx
);

Fails, with a counterexample

2 Refine

assert property (
  commit_valid |->
  !commit_older_pending
);

Passes. A stronger assertion

3 Select

Best assertion, chosen by the search

4 Learn

Counterexample and the lesson, stored for the next design

Continual learning. The stored lesson comes back into the next generation.

Early results

Measured, with the limits written down.

  • 7 of 9 vs 3 of 10Open-source bug cases found, against one plain prompt to the same model. Provisional: one seed for Chipster, two for the prompt. Chipster also raised more false leads per bug (4.0 vs 2.5).
  • PR #81On Xilinx open-nic-shell, one of 29 leads matched a known open bug. A directed test reproduces it. The other 28 were environment or tool issues.
  • 1,010 assertionsOn an open I2C controller: 558 proved, 315 with counterexamples. An earlier run. Counterexamples are leads for review, not automatic RTL bugs.

Bring us a block.

Chipster is in early access.