XaveT :
If you add a second chip to check the first chip's calculations... isn't that second chip performing the work of the first one too? Or a second unit in the same chip? So why use the first chip/unit at all? Plus, who's to say that the second chip won't be modified if that's what they are worried about?
This seems like a monumental waste of research resources to me.
Read more carefully, and check the Source link (below the headline), when you have questions. The verification module is simpler than the actual chip under test, and you don't need many of them. So, you can fab it at a trusted foundry and use it to test your production wafers. They needn't be as fast, either, so you can fab them on an older, cheaper process.
Honestly, I think it's a good problem to solve. I think it's rare that this would really happen, but I expect sometimes it might.
If you want to understand how it actually
works, here's the paper:
https://eprint.iacr.org/2015/1243.pdf