By: wumpus (wumpus.delete@this.lost.in.a.hole), March 22, 2021 8:33 am
Room: Moderated Discussions
Linus Torvalds (torvalds.delete@this.linux-foundation.org) on March 21, 2021 2:38 pm wrote:
> wumpus (wumpus.delete.delete@this.this.lost.in.a.hole) on March 21, 2021 12:24 pm wrote:
> >
> > Your best (only) bet would be to forget about a new architecture for general purpose programming
> > and try to optimize some niche with a market big enough to support it. Best guess is machine learning,
>
> I think that's true, but it's a bit of a cop-out. It's not replacing - or even really
> modifying - the core CPU architecture, it's just adding logic on the side.
>
The idea was that the the "next architecture" would be a general purpose processor. But on second thought, the idea of any "next general purpose processor" outcompeting GPUs and even more specifically designed chips appears impossible (and I admitted it would wind up being customized specifically for ML).
My "or possibly SQL" was originally a throwaway statement to point out that there was nothing specific to ML other than a wide open market. But SQL doesn't appear to have better accelerators than a strong general purpose processor, and a huge number of expensive computers are bought to run nothing (significantly) else. So if you made a sufficiently superior general purpose processor, your clear market should be databases in general and SQL in particular.
I'm more or less blanking on anything else that is small enough to build enough of "everything else" and big enough to support an architecture.
So my suggestion is perhaps it isn't necessary to replace all at once the trillion dollars of infrastructure that goes into more or less general purpose processors designed to execute sequential code. Perhaps all you need is to get your "next generation processor" to significantly improve the performance of SQL and get it in the server room/datacenter.
It would almost be the inverse of your "key for ARM on the desktop": start with servers and let developers use retired servers as development/experimental boxes, but it is means of getting a "blue sky" architecture from concept to total world domination. You don't have to be everything to everyone for anything but the dominant architecture. But you have to give enough people enough reason to jump ship to keep an architecture going.
> wumpus (wumpus.delete.delete@this.this.lost.in.a.hole) on March 21, 2021 12:24 pm wrote:
> >
> > Your best (only) bet would be to forget about a new architecture for general purpose programming
> > and try to optimize some niche with a market big enough to support it. Best guess is machine learning,
>
> I think that's true, but it's a bit of a cop-out. It's not replacing - or even really
> modifying - the core CPU architecture, it's just adding logic on the side.
>
The idea was that the the "next architecture" would be a general purpose processor. But on second thought, the idea of any "next general purpose processor" outcompeting GPUs and even more specifically designed chips appears impossible (and I admitted it would wind up being customized specifically for ML).
My "or possibly SQL" was originally a throwaway statement to point out that there was nothing specific to ML other than a wide open market. But SQL doesn't appear to have better accelerators than a strong general purpose processor, and a huge number of expensive computers are bought to run nothing (significantly) else. So if you made a sufficiently superior general purpose processor, your clear market should be databases in general and SQL in particular.
I'm more or less blanking on anything else that is small enough to build enough of "everything else" and big enough to support an architecture.
So my suggestion is perhaps it isn't necessary to replace all at once the trillion dollars of infrastructure that goes into more or less general purpose processors designed to execute sequential code. Perhaps all you need is to get your "next generation processor" to significantly improve the performance of SQL and get it in the server room/datacenter.
It would almost be the inverse of your "key for ARM on the desktop": start with servers and let developers use retired servers as development/experimental boxes, but it is means of getting a "blue sky" architecture from concept to total world domination. You don't have to be everything to everyone for anything but the dominant architecture. But you have to give enough people enough reason to jump ship to keep an architecture going.
