First resultMXCore

Arithmetic is the design choice.

pre spin changes the datapath arithmetic at RTL and proves the change with a formal error bound.

Rim, turning half a turn. The cut stays level at every angle. Drawn in code

Changed the arithmetic. Timing still met. No measurable perplexity change.

How it works

RTL in, RTL out. In between, a bound you can check.

Your RTL error bound formal, at RTL Python model Optimised RTL Test vectors Your RTL error bound formal, at RTL Python model Optimised RTL Test vectors
Diagram, not data.
  1. 01

    Your RTL

    Send the datapath as RTL. A throwaway block is a fine place to start.

  2. 02

    A formal bound

    We change the arithmetic at RTL and prove the change with a formal error bound.

  3. 03

    Back to you

    Optimised RTL, a bit-exact Python model and test vectors. The error bound is yours to check.

The result

Timing met. Area and power down.

MXCore, patched against its own baseline. One design so far.
MeasureChangeScope
Area−9.7%Whole MXCore
Power−6.0%MXFP8 engine
Perplexity−0.005%No measurable change
TimingMetMXCore
Figure 1 Illustration, not measured data. One MXFP8 dot-product unit: 32 E4M3 results from baseline above, patched unit below. The few low mantissa bits that differ are shown in blue.

N=1.

One design so far, MXCore. Yours could be N=2.

Notes

  • Where the error bound comes from

    MethodPlaceholder title

  • Bit-exact, and how we check it

    NotesPlaceholder title

  • MXCore, before and after

    ResultPlaceholder title

Team

“We” is two people.

Start with a throwaway block of RTL.

Or a conversation about your flow. Confidentiality can be agreed upfront.