Interesting read. Still no information how it compares to other HSM implementations however, and why they didn't use the more well studied PKCS protocols in theirs.
I would imagine the problem was more with existing implementations rather than some problems with HSMs per se, as they probably have much larger latency than you normally have so things like blocking reads start to matter.
I guess you mean PKCS#11, but that isn't a protocol. It's a C API. Trying to funnel it over a network would be both extremely complex, and non-standard.
No, I was thinking of PKCS#7 which you can run over IP if you want. But it's a complicated set of standards and I'm not sure how they inter-relate. That's why I summarized them as PKCS-style protocols.
This is not some theoretical thing. You can buy these devices off the shelf. If you have worked with PKI you have seen them, or some variant thereof.
I would imagine the problem was more with existing implementations rather than some problems with HSMs per se, as they probably have much larger latency than you normally have so things like blocking reads start to matter.