Email the founders

00 RTL datapath arithmetic

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.

In
Your RTL
Check
Formal error bound
Out
Optimised RTL, a bit-exact Python model, test vectors

01 Method

RTL in. A bound you can check. RTL out.

How pre spin works Your RTL goes in. pre spin changes the datapath arithmetic and proves the change with a formal error bound. Out come optimised RTL, a bit-exact Python model and test vectors. IN Your RTL datapath block PRE SPIN Arithmetic changed at RTL + bound − bound Formal error bound OUT 1 Optimised RTL OUT 2 Bit-exact Python model OUT 3 Test vectors How pre spin works: your RTL in, a formal error bound, then optimised RTL, a bit-exact Python model and test vectors out. IN Your RTL datapath block PRE SPIN Arithmetic changed at RTL Formal error bound OUT 1 Optimised RTL OUT 2 Bit-exact Python model OUT 3 Test vectors
  1. In

    A block of your RTL. A throwaway one is fine to start.

  2. Change

    pre spin changes the datapath arithmetic at RTL, with a formal error bound on the result.

  3. Out

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

02 Result, MXCore

Timing met. Area and power down.

MXCore, baseline against the unit pre spin returned
MeasureScopeChange
AreaWhole MXCore−9.7%
PowerMXFP8 engine−6.0%
PerplexityNo measurable change−0.005%
TimingMXCoreMet
Baseline S E E E E M M M Patched S E E E E M M M Results 1 to 16 Baseline S E E E E M M M Patched S E E E E M M M Results 17 to 32 Baseline S E E E E M M M Patched S E E E E M M M
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 colour.

03 N=1

One design. Yours next.

One design so far: MXCore. Start with a throwaway block of RTL or a conversation about your flow. Confidentiality can be agreed upfront.

05 Founders

Two people, one inbox each

Email the founders

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