Arithmetic is the design choice.
Abstract. 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.
Baseline32 × E4M3
Patched9 bits differ
Each column is one result. Rows, top to bottom: sign, four exponent bits, three mantissa bits.
Baseline32 × E4M3
Patched9 differ
Each row is one result. Bits, left to right: sign, four exponent bits, three mantissa bits.
RTL in, formal error bound, optimised RTL out
If you own the datapath in an inference accelerator, a failed block is on you, so you discount vendor numbers you cannot check. That is the right instinct. Everything below is written so that you can check it.
The usual optimisation suites work around the arithmetic in a datapath and leave it as it was written. pre spin changes the arithmetic itself, at RTL, and proves the change.
The proof is a formal error bound for the change. With it come a bit-exact Python model of the patched unit and test vectors, checked against simulation of the patched RTL. You do not have to take our numbers on trust. The error bound is yours to check.
- In Your datapath RTL
- Change The arithmetic, with a formal error bound
- Out Optimised RTL, a bit-exact Python model and test vectors
Timing met. Area and power down.
Changed the arithmetic. Timing still met. No measurable perplexity change. Every figure below is measured on MXCoreb against its own baseline, and each one carries its scope.
Area is measured over the whole of MXCore. Power is measured on its 4-unit MXFP8 engine, not on the whole accelerator; that scope is narrower, and the table says so. Perplexity was measured on Qwen2.5-7B in MXFP8, where −0.005% is no measurable change.
One design is one design. We have not shown that the method carries over to every datapath, and this page does not claim it.
| Measure | Change | Scope |
|---|---|---|
| Timing | Met | Patched MXCore, after the change |
| Area | −9.7% | Whole MXCore |
| Power | −6.0% | 4-unit MXFP8 engine, not the whole accelerator |
| Perplexity | −0.005% | Qwen2.5-7B, MXFP8. No measurable change |
Two people
When this page says we, it means the two of us.
- Tom de Vriezetom@prespin.com
- Leo Strijbosleo@prespin.com