I don't think he knows any, he didn't mention any, and he's probably just pontificating on general principles.
I would say the whole point of RISC-V is to study the mistakes of past ISAs and avoid them. The process for proposing and ratifying new extensions involves very wide consultation with industry and academic experts on particular topics e.g. memory model, vector processing, cache control instructions, hypervisors, etc as opposed to Intel's or Arm's internal development of features and then "Here it is, hope you'll like it!" unexpected announcement.
The RISC-V setup certainly allows -- even encourages -- individual vendors (e.g. THead, Andes, LowRISC & ETH Zurich with CHERI) to make experiments and make their own mistakes. That's competition: on features, on time to market, on styles of doing things.
Successful experiments may be standardised in due course.
The main "mistake" so far has been standardisation more careful and taking longer than some people would like, leading them to shipping draft versions (RVV 0.7.1, and a guess where the `VS` bits would go in `mstatus`) or just best guesses (THead's PMA) because their customers needed them.
6 comments
[ 2.2 ms ] story [ 25.6 ms ] threadLinus Torvalds on security, AI, open source and trust- https://www.youtube.com/watch?v=cPvRIWXNgaM
I would say the whole point of RISC-V is to study the mistakes of past ISAs and avoid them. The process for proposing and ratifying new extensions involves very wide consultation with industry and academic experts on particular topics e.g. memory model, vector processing, cache control instructions, hypervisors, etc as opposed to Intel's or Arm's internal development of features and then "Here it is, hope you'll like it!" unexpected announcement.
The RISC-V setup certainly allows -- even encourages -- individual vendors (e.g. THead, Andes, LowRISC & ETH Zurich with CHERI) to make experiments and make their own mistakes. That's competition: on features, on time to market, on styles of doing things.
Successful experiments may be standardised in due course.
The main "mistake" so far has been standardisation more careful and taking longer than some people would like, leading them to shipping draft versions (RVV 0.7.1, and a guess where the `VS` bits would go in `mstatus`) or just best guesses (THead's PMA) because their customers needed them.
But they are not avoiding them.