How did you find developing this data parallel R1CS for VMs? like a whole processor.

Yeah, I think maybe if we take a step back, if you look at how existing ZKVMs work prior to Magnet, I think there were broadly two approaches to designing ZKVMs.

On the one side, we have the hash and more ZKVMs.

All of them, they organize their execution across multiple witness vectors.

So there is a separate witness vector, or I think it's also commonly referred to as a table.

So you have separate tables.

And then each table is just responsible for one instruction.

But why are you getting this improvement? Where is that coming from? You talked about the scheme and how you could plug and play stuff, but I don't really get what part of the stack is leading to that efficiency gain.