Description
User story: "As a reader of the report, I'd like to know the source of the discrepancy between the two implementations, so that I can understand to what extent errors are being introduced on different systems."
In his response to the first version of the report, Rich wrote the following:
Do we have evidence for the source of the discrepancy for the two implementations. Is it something as basic as matrix-matrix multiplication being implemented slightly differently (e.g. different orders of sum/product ops)? Or is there another source? It would be good to know so we can predict when this is going to be an issue. E.g. have you compared basic ops between the two systems?
To try to answer this we should run some tests, performing basic matrix calculations at different precisions and compare the results carefully.