Comment by hajile
7 hours ago
You can choose to develop proprietary extensions, but who’s going to use them?
A great case study is the companies that implemented the pre-release vector standard in their chips.
The final version is different in a few key ways. Despite substantial similarities to the ratified version, very few people are coding SIMD for those chips.
If a proprietary extension does something actually useful to everyone, it’ll either be turned into an open standard or a new open standard will be created to replace it. In either case, it isn’t an issue.
The only place I see proprietary extensions surviving is in the embedded space where they already do this kind of stuff, but even that seems to be the exception with the RISCV chips I’ve seen. Using standard compilers and tooling instead of a crappy custom toolchain (probably built on an old version of Eclipse) is just nicer (And cheaper for chip makers).
Yes, extensions are perfect for embedded. But not just there.
Extensions allow you to address specific customer needs, evolve specific use cases, and experiment. AI is another perfect fit. And the hyperscaler market is another one where the hardware and software may come from the same party and be designed to work together. Compatibility with the standard is great for toolchains and off-the-shelf software but there is no need for a hyperscaler or AI specific extension to be implemented by anybody else. If something more universally useful is discovered by one party, it can be added to a future standard profile.