AVX-512 What software use AVX-512?

[rwnrwnn7]

AVX-512 What software use AVX-512?

this is a new set of commands from Intel
1. which programs use this command set?

 

[The Stilt]

Latest builds of X264 have some preliminary support in them.
5-10% performance improvement over AVX2, depending on the used preset currently (according to the devs).

 

[rwnrwnn7]

thanks - what software?
will I see this on KLITE?

 

[TheGiant]

Latest builds of X264 have some preliminary support in them.
5-10% performance improvement over AVX2, depending on the used preset currently (according to the devs).

from your all around deep knowledge what do you think is the realistic advantage of AVX512 over AVX2?

thanks

 

[The Stilt]

thanks - what software?
will I see this on KLITE?

X264 encoder.
http://git.videolan.org/?p=x264.git;a=summary

 

[rwnrwnn7]

thanks
Is AMD allowed to use these commands ( 512 ) ?
Or do they have similar / another name?

 

[rwnrwnn7]

thanks
Is AMD allowed to use these commands ( 512 ) ?
Or do they have similar / another name?

 

[The Stilt]

thanks
Is AMD allowed to use these commands ( 512 ) ?
Or do they have similar / another name?

In theory they could add support for AVX512, however there isn't any performance benefit from doing so (more likely a small penalty, similar to AVX2) without increasing the resources in the core.

 

[rwnrwnn7]

increasing the resources in the core.

I did not understand what " resource "
in theory AVX-512 should bring 100% over AVX2 ?

 

[Dufus]

IIRC Firestarter can use AVX512 but seemed a little buggy last time I tried it.

 

[Topweasel]

I did not understand what " resource "
in theory AVX-512 should bring 100% over AVX2 ?

You have to have dedicated pipelines for processing this kind of work and basically an answer sheet.

SkyLake-X includes only a fraction of AVX-512 (which is another problem with AVX-512) and the "answer sheet" is alone the size of a whole Atom core. AVX-512 also takes up close to 20% of each core. This is for functionality almost non-existant in the open world meant for HPC and other super high end processing.

Add into the effect that AVX2 has over AVX and AVX 512 over AVX2 in power and temperature (and clock speeds without the offset) and you can see why AMD didn't even do AVX2 straight up and went for combining AVX units for AVX2.

That is the resource usage. Tons of space, increased power usage, increased temps, lower clocks.

 

[tamz_msc]

I did not understand what " resource "
in theory AVX-512 should bring 100% over AVX2 ?

There is supposed to be no difference in theory and practice, but in practice there is. 100% more performance over AVX2 would only happen ideally in something like LINPACK.

AVX512 advantages over AVX2 in practically useful work is extremely limited in both applicability and performance.

 

[NostaSeronx]

AVX-512 is mostly about increasing efficiency and enhancing the amount of workloads that can be vectorized.

https://dl.acm.org/citation.cfm?id=2908111
Also, AVX-512 already has a bunch of patches...

https://colfaxresearch.com/skl-avx512/
^-- plenty of new stuff that would have to be worked into existing apps.

So, only expect AVX-512 to become rampantly relevant when it becomes mainstream and is in like Celerons/Pentiums.

 

[PhonakV30]

What's point of AVX2/AVX512 when We do have GPU workloads? Isn't GPU rival to AVX?

 

[jpiniero]

What's point of AVX2/AVX512 when We do have GPU workloads? Isn't GPU rival to AVX?

Pretty much, but that is sort of why Intel did it. They did manage to get a couple HPC deals.

 

[Carfax83]

AVX512 advantages over AVX2 in practically useful work is extremely limited in both applicability and performance.

AVX512VL should allow AVX512 to be much more practical and useful in general computing.

Of course this will take years to manifest unfortunately.

 

[tamz_msc]

AVX512VL should allow AVX512 to be much more practical and useful in general computing.

Of course this will take years to manifest unfortunately.

AVX512 Integer performance is still only 20 (or 40, can't remember) percent over AVX2, according to Sandra. As for FP, it's still limited to array or matrix operations as long as they don't run into memory bottlenecks, and in isolation. Fawning over big performance deltas in MKL FFT/LINPACK isn't really what you'd call an example of overall usefulness. When software is written with varying degrees of competence when it comes to exposing vectorizable paths, performance is going to be all over the place. And I'm not even talking about the physical restrictions of die area, clock speeds and power consumption.

By the time 512 bit wide vector SIMD becomes useful, CPUs will already have more 128 or 256 bit execution units in the same die area to tackle the problem.

 

[IRobot23]

Latest builds of X264 have some preliminary support in them.
5-10% performance improvement over AVX2, depending on the used preset currently (according to the devs).

If I may ask you. How much more power does it needs?

 

[Topweasel]

I doubt anyone outside Intel knows what the exact power difference is between them but AVX 512 on SL-X requires something close to a 500MHz reduction in clocks to keep TDP in line when running the code. AVX2 on KL needed something like a 300MHz offset.

 

[The Stilt]

If I may ask you. How much more power does it needs?

No idea.

 

[Eric1987]

X264 encoder.
http://git.videolan.org/?p=x264.git;a=summary

Why is 264 still being developed? 265 is superior in every way last time I checked.

 

[eddman]

Why is 264 still being developed? 265 is superior in every way last time I checked.

There are many devices that cannot do full HEVC/h.265 hardware decoding, be it CPUs, GPUs, mobile phones, set-top boxes, etc. Companies cannot abandon h.264 just yet, until the large majority of people's devices support h.265.

P.S. This is off-topic though.

 

[Eric1987]

There are many devices that cannot do full HEVC/h.265 hardware decoding, be it CPUs, GPUs, mobile phones, set-top boxes, etc. Companies cannot abandon h.264 just yet, until the large majority of people's devices support h.265.

P.S. This is off-topic though.

Sorry I just really didn't know. Thought it was something I was missing and it was.

 

[Jan Olšan]

x264 instances probably still power vast majority of streaming video on the internet.

 

[DeathReborn]

The new Time Spy DX12 benchmark will make use of AVX-512, don't know if that will help or not but will be interesting to see.

https://www.anandtech.com/show/1189...remarks-time-spy-extreme-dx12-benchmark-in-4k