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.
RTL in, formal error bound, optimised RTL out.
- In Your RTL A datapath unit, as you wrote it.
- 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.
-
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.
The error bound is yours to check.
The result
Changed the arithmetic. Timing still met. No measurable perplexity change.
| Measure | Scope | Change |
|---|---|---|
| Area | Whole MXCore | −9.7% |
| Power | 4-unit MXFP8 engine, not the whole accelerator | −6.0% |
| Perplexity | Qwen2.5-7B, MXFP8. No measurable change. | −0.005% |
| Timing | MXCore | Met |
One design so far, MXCore. Yours could be N=2.
Notes
Dated notes on the work, as it happens. The titles below are placeholders.
- 01Reading a formal error boundPlaceholder
- 02MXCore, baseline against patchedPlaceholder
- 03What the bit-exact model is forPlaceholder
Founders
- Tom de Vriezetom@prespin.com
- Leo Strijbosleo@prespin.com
One design. Yours next.
Start with a throwaway block of RTL or a conversation about your flow. Confidentiality can be agreed upfront.