Verifiable Matrix Multiplication for Consensus Mechanisms

Been diving into a consensus mechanism where the core task is verifiable matrix multiplication. What's fascinating is how we can reconstruct a precise floating-point product using INT8 arithmetic. Sounds wild, right? It's kinda the same thing ML does during its lossy quantization process. So here’s the deal with the core technique, called the Ozaki Scheme. It’s all about error-free transformations. The plan is to completely skip rounding errors. You break each floating-point matrix into a combination of bounded integer slices, all fitting snugly within signed INT8. Then you multiply these slice pairs... I think this could change the game for how we think about compute layers in consensus.

+1, this is super interesting!

Not sure if this can be done, honestly. INT8 has limitations.

@user1, I think they’re onto something. This method could really help avoid those pesky rounding issues.

So how exactly do these integer slices work? Like, can you give a bit more detail?

Totally. The idea is that by splitting the matrices this way, you can manage the errors better and still get reliable results.

Sounds complicated... but I’m intrigued.

Pruфы? Has this been tested anywhere?

Not yet, but I’m working on some examples to showcase the process. Stay tuned!