Email the founders

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

Measured on MXCore

Area
−9.7%
Power
−6.0%
Timing
Met

How it works

RTL in, formal error bound, optimised RTL out.

  1. 01Input

    Your RTL

    The datapath block as it stands. Start with a throwaway block if you prefer.

    Format
    RTL
    Confidentiality
    Can be agreed upfront
  2. 02Change

    A formal error bound

    The datapath arithmetic changes, and a formal error bound proves the change.

    Bound
    Formal
    Level
    RTL
  3. 03Output

    Three things back

    Optimised RTL, a bit-exact Python model and test vectors, so you can check it yourself.

    Optimised RTL
    Yes
    Python model
    Bit-exact
    Test vectors
    Yes

The result

Timing met. Area and power down.

MeasureScopeChange
AreaWhole MXCore−9.7%
PowerMXFP8 engine−6.0%
PerplexityNo measurable change−0.005%
TimingMXCoreMet

MXCore, patched against baseline. One design so far.

Baseline
Patched
Baseline
Patched
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 .

One design. Yours next.

N = 1 MXCore is the only design so far. We say so first.

Notes

Notes from the bench. Titles are placeholders until the posts are written.

Team

Two founders. “We” on this page means Tom and Leo.

The error bound is yours to check.

Email the founders

Start with a throwaway block of RTL or a conversation about your flow. Confidentiality can be agreed upfront.