Question What will going to happen when AVX1024 is launched by Intel?

Tugrul_8192bit

Junior Member
Apr 22, 2015
13
2
81
I guess it will have even lower boost frequencies but much better performance per core, especially with 1024 bit register (64 of them = 64kbit / 8kB fastest memory, maybe with 500-1000 GB/s bandwidth per core).
 
Last edited:
Jul 27, 2020
16,237
10,288
106
  • Like
Reactions: Tlh97 and Schmide

zir_blazer

Golden Member
Jun 6, 2013
1,164
406
136
You mean Intel AMX (Advanced Matrix Extensions)? Regular users most likely aren't even going to get them at all, more so given how Intel recently put in limbo other existing extensions like AVX-512, TSX and SGX.
If AMX actually gets implemented, you will most likely see the same thing than we already saw with AVX-512: Even lower clocks than AVX2, but higher performance for optimized code. Or were you expecting anything revolutionary?
 

DrMrLordX

Lifer
Apr 27, 2000
21,626
10,841
136
If Intel has any common sense, they'll form a SIMD consortium with ARMand/or Fujitsu to try and open up SVE2 or a future version (SVE2.x, SVE3, whatever).
 

soresu

Platinum Member
Dec 19, 2014
2,657
1,860
136
You mean Intel AMX (Advanced Matrix Extensions)? Regular users most likely aren't even going to get them at all, more so given how Intel recently put in limbo other existing extensions like AVX-512, TSX and SGX.
If AMX actually gets implemented, you will most likely see the same thing than we already saw with AVX-512: Even lower clocks than AVX2, but higher performance for optimized code. Or were you expecting anything revolutionary?
AMX is targeted very specifically at AI/ML type operations on the CPU.

As they (until more recently) lacked any significant GPU power of their own to do those kind of operations it was an obvious next step to create such an AI/ML focused extension just as ARM did to spread AI/ML ops across its entire IP family.

As you say though it will likely have little use precisely because GPUs and ML focused accelerator cards can be drastically more efficient in this use case.
 

diediealldie

Member
May 9, 2020
77
68
61
It won't have any impacts unless memory subsystems are improved substantially. 128B per cycle means 2 L1D cache lines and 2 cycles of backend to L1D store bandwidth .
 

inf64

Diamond Member
Mar 11, 2011
3,698
4,018
136
I've been around, but I just don't find much to say at the moment.

I'm also actually all in on AMD too, as I have an AMD laptop that I am very happy with and have an AM4 system that I will be keeping for some time.
I'm in the same boat. I plan on buying 5800X3D when the price drops some more in the coming months. For laptop, I plan on buying 6800U based model early next year.