4 ms·

All I see is a compatibility mess. Intel got their money from corporations and large servers. If this thing doesn't play well with old software using hyperthrea

All I see is a compatibility mess. Intel got their money from corporations and large servers. If this thing doesn't play well with old software using hyperthreading logic, shit will break in a million and one ways.

sydbarrett74 3y ago
Since the patent describes the scheduling hardware as sending 'suggestions' to the OS, it stands to reason that the OS would be free to ignore these suggestions and operate conventionally. Of course nobody will know until Microsoft and other vendors have sample chips to play with.
wmf 3y ago
Software mostly isn't aware of SMT anyway and some processors never had it so software will have no trouble.
tommiegannert 3y ago
Looks like it's just about collecting stats on whether code "partitions" executed the way the (new) predictor thought it would, and if it didn't it, suggests to the OS to move the load to another type of core. I doubt this would break anything except possibly performance regressions in tightly optimized code that's been tailored to very specific CPU HT behavior. From a cursory glance, it seems like an obvious thing to do now that there are cores of different types (Efficiency and Performance). Just like deciding at which level to cache something, when caches became hierarchies. Edit: BTW, when we have FPGAs as another type of "core", it'll be cool to see if the OS can decide to not only schedule code on that subsystem, but also select a completely different instruction set ("bit stream") for that section of code. Would require a true "multi-architecture" binary, with the OS deciding "architecture" for individual code partitions depending on what other processes are running on the machine.
- wmf 3y ago
  when we have FPGAs as another type of "core" This isn't going to happen because FPGAs are too area-inefficient.
pjmlp 3y ago
I have disabled it in several occasions as it was only a stop game measure for high core count, and performance isn't that great.

sydbarrett74 3y ago

Since the patent describes the scheduling hardware as sending 'suggestions' to the OS, it stands to reason that the OS would be free to ignore these suggestions and operate conventionally. Of course nobody will know until Microsoft and other vendors have sample chips to play with.

wmf 3y ago

Software mostly isn't aware of SMT anyway and some processors never had it so software will have no trouble.

tommiegannert 3y ago

Looks like it's just about collecting stats on whether code "partitions" executed the way the (new) predictor thought it would, and if it didn't it, suggests to the OS to move the load to another type of core. I doubt this would break anything except possibly performance regressions in tightly optimized code that's been tailored to very specific CPU HT behavior. From a cursory glance, it seems like an obvious thing to do now that there are cores of different types (Efficiency and Performance). Just like deciding at which level to cache something, when caches became hierarchies. Edit: BTW, when we have FPGAs as another type of "core", it'll be cool to see if the OS can decide to not only schedule code on that subsystem, but also select a completely different instruction set ("bit stream") for that section of code. Would require a true "multi-architecture" binary, with the OS deciding "architecture" for individual code partitions depending on what other processes are running on the machine.

wmf 3y ago
when we have FPGAs as another type of "core" This isn't going to happen because FPGAs are too area-inefficient.

pjmlp 3y ago

I have disabled it in several occasions as it was only a stop game measure for high core count, and performance isn't that great.