News Intel's Clear Linux distribution adds support for APX and AVX10, which aren't out yet

Status
Not open for further replies.
This is a nothing burger. AVX10.1 adds no practical benefits over AVX-512, at the level Granite Rapids will support the latter.

Furthermore, Intel hasn't announced APX support in any CPU set to launch this year. I think the APX aspect is probably just so that Intel can start doing internal testing, profiling, & tuning so that compiler support for APX is in good shape, by the time supporting silicon finally does launch.
 
In the future, this implementation of AVX10 will allow AVX-512 applications to run across both Intel's P-Cores and E-Cores, among other improvements.
No, it won't.

AVX10 will enable client CPUs to use AVX-512 features and semantics, but only at up to 256-bit widths. Their docs refer to this as AVX10/256.

Intel APX, meanwhile, ... can also serve as a method of boosting AI performance, according to Intel's documentation.
No, it doesn't.

You read that too quickly. The only place the linked document mentions AI performance is in reference to AMX, which is completely different and unrelated.

Contrary to all these vector & matrix extensions, the interesting thing about APX is that it should improve performance on general-purpose computation!


C'mon, guys. I know this stuff is complex, but these details actually matter!
 
Last edited:
  • Like
Reactions: Order 66
Nothing of the sort has been announced. AVX10.2 does not imply 512-bit.

I'm not saying 512-bit will never return to Intel's client CPUs, but the first thing we should see in them is AVX10/256.
This is definitely right (10.2 is AVX10/256 for client) and I'd be surprised if that changed any time soon.
I think the APX aspect is probably just so that Intel can start doing internal testing, profiling, & tuning so that compiler support for APX is in good shape, by the time supporting silicon finally does launch.
I'd bet that's the same reasoning for AVX10 as well as I'm sure Intel wants to make recompiling to 256 as painless as possible. They seem to do an awful lot of really good customization work with Clear so it makes sense to me that they'd get everything in as early as possible.

Perhaps for E-cores Intel will eventually go the route AMD has for AVX512 implementation as it seems to be the most efficient way to go about it. I'm just not sure it'll be worth the die space unless there are calls for it from their high core count customers.
 
Status
Not open for further replies.