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.

baseline s e e e e m m m patched s e e e e m m m 0 8 16 24 31 baseline s e e e e m m m patched s e e e e m m m 0 4 8 12 16 20 24 28 31
Fig. 1 Illustration, not measured data. One MXFP8 dot-product unit: 32 E4M3 results from baseline aboveon the left, patched unit belowon the right. The few low mantissa bits that differ are shown in colour.

How it works

You send datapath RTL. pre spin changes its arithmetic and states a formal error bound for the change. You get back optimised RTL, a bit-exact Python model and test vectors, checked against simulation of the patched RTL.

RTL in, bound, RTL outFig. 2
  1. 01In Datapath RTL.
  2. 02Change The arithmetic, with a formal error bound.
  3. 03Out Optimised RTL, a bit-exact Python model and test vectors.

Timing met. Area and power down.

Measured on MXCore. Changed the arithmetic. Timing still met. No measurable perplexity change.

MXCore, against its baselineTable 1
MXCore results against the baseline, with the scope of each figure.
#MeasureScopeChange, drawn to scaleValue
1 Area whole MXCore −9.7%
2 Power 4-unit MXFP8 engine, not whole accelerator −6.0%
3 Perplexity Qwen2.5-7B, MXFP8; no measurable change −0.005%
4 Timing MXCore met

Bars to scale, 0 to 10%. At this scale 0.005% is narrower than a pixel.

One design. Yours next.

One design so far, MXCore.

Founders

pre spin is two people. Write to either of us.

Email the founders

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