BaselineSEEEEMMMPatchedSEEEEMMM132signexponentmantissalow bits differ
Figure 1. Illustration, not measured data. One MXFP8 dot-product unit: 32 E4M3 resultsthe first 16 of its 32 E4M3 results from baseline above, patched unit below. The few low mantissa bits that differ are shown in .

Arithmetic is the design choice.

pre spin changes the datapath arithmetic at RTL and proves the change with a formal error bound. One design so far. No measurable perplexity change.

Email the founders See the result

RTL in, formal error bound, optimised RTL out.

  1. In Your RTL A datapath unit, as you wrote it.
  2. pre spin Changes the arithmetic and proves the change with a formal error bound: a stated limit on the most it can move a result.
  3. Out
    • Optimised RTLYour unit, with its arithmetic changed.
    • A bit-exact Python modelMatches the patched RTL bit for bit.
    • Test vectorsChecked against simulation of the patched RTL.
Figure 2. What goes in, and the three things that come back.

The error bound is yours to check.

The result

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

Table 1. Timing met. Area and power down. MXCore, baseline against patched dot-product units.
MeasureScopeChange
AreaWhole MXCore−9.7%
Power4-unit MXFP8 engine, not the whole accelerator−6.0%
PerplexityQwen2.5-7B, MXFP8. No measurable change.−0.005%
TimingMXCoreMet

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

Notes

Dated notes on the work, as it happens. The titles below are placeholders.

  1. 01Reading a formal error boundPlaceholder
  2. 02MXCore, baseline against patchedPlaceholder
  3. 03What the bit-exact model is forPlaceholder

Founders

One design. Yours next.

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

Email the founders