Question The AVX-512 thread

[igor_kavinski]

Here's your chance to spam a thread with all manner of AVX-512 related stuff, including

technical discussion, benchmarks,

software including libraries explicitly designed to take advantage of AVX-512 acceleration,

your personal experiences enjoying AVX-512 acceleration,

your future hardware/software wish list with AVX-512 support,

if you dropped the big bucks on Threadripper or server CPUs mainly for AVX-512 etc.


I'll start off with some links:

The origins of AVX-512: https://tomforsyth1000.github.io/papers/LRBNI origins v4 full fat.pdf

Benchmarking the performance and energy consumption of the AVX512 and VNNI instruction sets: https://addi.ehu.es/bitstream/handle/10810/58088/TFG_Jon_Arriaran.pdf?sequence=2

Using the most unhinged AVX-512 instruction to make the fastest phrase search algo

Disclaimers before we start For those who don’t want to read/don’t care that much, here are the results. I hope after seeing them you are compelled to read. TL;DR: I wrote a super fast phrase search algorithm using AVX-512 and achieved wins up to 1600x the performance of Meilisearch. The source...

gab-menezes.github.io

Caching Energy Efficiency Data – Mobile and AVX-512

A few days ago, we looked at the power costs associated with moving data on some desktop parts.

chipsandcheese.com

https://albertvilella.substack.com/p/intels-avx-512-use-cases-part1 (sadly requires subscription for full read)

Fabian Giesen (@rygorous@mastodon.gamedev.place)

By request, my usual "the least interesting part about AVX-512 is the 512 bits vector width" infodump in thread form. So here goes, a laundry list of things introduced with AVX-512 that I think are way more important to typical use cases than the 512-bit vectors are:

mastodon.gamedev.place

Optimize for Intel® AVX-512 with or without Intel® AVX-512 hardware

Optimizing vectorization becomes increasingly important as new technology like AVX-512 becomes available. Software must be both efficiently threaded and efficiently vectorized to fully utilize modern hardware.

www.intel.com

GitHub - twest820/AVX-512: AVX-512 documentation beyond what Intel provides

AVX-512 documentation beyond what Intel provides. Contribute to twest820/AVX-512 development by creating an account on GitHub.

github.com

Zen5's AVX512 Teardown + More...

Transcoding Unicode with AVX-512: AMD Zen 4 vs. Intel Ice Lake – Daniel Lemire's blog

lemire.me

An Intro to AVX-512 Assembly Programming

Brief history of AVX-512, interesting feature sets, how-to tutorials and an introduction to assembly programming using Intel AVX-512 instructions.

www.physicsforums.com

Parallel Programming on a CPU with AVX-512

In this article, I describe a program that uses Intel AVX-512 assembly instructions and includes a comparison of the results from both programs.

www.physicsforums.com

AVX-512 Programming: Extracting Column Subtotals from a Table

In this Insights article I'll present an example that shows how Intel® AVX-512 instructions can be used to read a whole row of data in a single operation, and then generate subtotals.

www.physicsforums.com

Vectorization Opportunities for Improved Performance with Intel® AVX-512

Examples of How Intel® Compilers Can Vectorize and Speed Up Loops

www.codeproject.com

AVX512/VBMI2: A Programmer’s Perspective

SingleStore is a modern relational database for cloud and on-premises that delivers immediate insights for modern applications and analytical systems. Book a demo or trial today!

www.singlestore.com

FFmpeg developers claim 94x speed boost using handwritten AVX-512 code

The FFmpeg team recently highlighted how AVX-512 instructions can deliver a significant performance boost in video processing workloads. According to a slide presented by one of the...

www.techspot.com

Latest AVX-512 Optimization For FFmpeg Shows Wild Improvement On AMD Ryzen - Phoronix

www.phoronix.com

https://www.reddit.com/r/hardware/comments/1ethe1e/quantifying_the_avx512_performance_impact_with

AMD Ryzen 7040 Series Shows Great AVX-512 Performance for Laptops / Mobile | Hacker News

news.ycombinator.com

 

[igor_kavinski]

Apparently Time Spy CPU test uses AVX-512? http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=threads/avx-512-what-software-use-avx-512.2520858/post-39110236

RobRich AVX-512 Chromium build: https://chromium.woolyss.com/

 

[Det0x]

Apparently Time Spy CPU test uses AVX-512? http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=threads/avx-512-what-software-use-avx-512.2520858/post-39110236

You need to enable/force AVX512 as a custom run.
Heres mine

 

[igor_kavinski]

You need to enable/force AVX512 as a custom run.
Heres mine
View attachment 118439

What's your normal score with just AVX2?

 

[Det0x]

What's your normal score with just AVX2?

Default settings

I scored 25 780 in Time Spy Extreme

AMD Ryzen 9 9950X, NVIDIA GeForce RTX 5090 x 1, 32768 MB, 64-bit Windows 11}

www.3dmark.com

But that is rank#2 cpu score in the world with a 9950X

 

[igor_kavinski]

+79% improvement????

Dude, you should try out Alan Wake 2 benchmark if AMD bios has option to enable/disable AVX-512. If these guys used AVX-512 in 3dmark, maybe they used it in their latest engine too?

 

[Markfw]

+79% improvement????

Dude, you should try out Alan Wake 2 benchmark if AMD bios has option to enable/disable AVX-512. If these guys used AVX-512 in 3dmark, maybe they used it in their latest engine too?

If you have an easy benchmark to run that uses a few cores and needs bandwidth, I can fir up my Turin 64 core. Its not as fast as a 9950x (2.3 ghz fully loaded) but 12 memory channels work really good if you need the bandwidth.

 

[igor_kavinski]

If you have an easy benchmark to run that uses a few cores and needs bandwidth, I can fir up my Turin 64 core. Its not as fast as a 9950x (2.3 ghz fully loaded) but 12 memory channels work really good if you need the bandwidth.

Maybe once I figure out how to run x265 encoder with and without AVX-512. I can then put it in the PES Handbrake benchmark package for easy testing.

 

[Markfw]

Maybe once I figure out how to run x265 encoder with and without AVX-512. I can then put it in the PES Handbrake benchmark package for easy testing.

even that doesn't seem easy. Running a prime95 task is easy. Stefan or Ken G6 are the experts at the hardest task in prime95.

My best guess is Seventeen or Bust LLR (SOB). It says very long tasks, but is it heavy in avx-512 ?

or maybe Sierpinski/Riesel Base 5 LLR (SR5) ????

 

[igor_kavinski]

DC tasks aren't easy to benchmark due to the variable work units they push to different clients. And you would have to disable AVX-512 in the BIOS if that is even possible if you want to see, suppose, whether the CPU crunches through more work units in 24 hours with AVX-512 than without.

 

[Markfw]

DC tasks aren't easy to benchmark due to the variable work units they push to different clients. And you would have to disable AVX-512 in the BIOS if that is even possible if you want to see, suppose, whether the CPU crunches through more work units in 24 hours with AVX-512 than without.

Well, disabling avx-512 is a possibility, but with these tasks, you will see upwords of 40% over Zen 4 and up to 80% over on Zen 5 vs NO avx-512.

It was just an idea that puts it in the best light.

 

[Nothingness]

A fascinating article about how Zen5 manages high frequency when running AVX-512 code. Interestingly it does IPC throttling to avoid getting too hot.

Zen 5's AVX-512 Frequency Behavior

Zen 5 is AMD's first core to use full-width AVX-512 datapaths.

chipsandcheese.com

 

[Jan Olšan]

Maybe once I figure out how to run x265 encoder with and without AVX-512. I can then put it in the PES Handbrake benchmark package for easy testing.

‑‑asm avx512 from commandline, for handbrake see picture in the article

Intel AVX-512 tested in x265: how to enable it and does it help? - HWCooling.net

AVX-512 for x265: how to use it, the effects on performance and the price you pay in power consumptionThe 11th generation Intel Core processors (Rocket Lake) are the first mainstream desktop parts supporting AVX-512 instruction set, previously only available in Xeons or the X299 platform. One...

www.hwcooling.net

 

[igor_kavinski]

‑‑asm avx512 from commandline, for handbrake see picture in the article

Unfortunately, it doesn't work for x265 which is what people are more interested in. I spent considerable time trying different variations of that command parameter with HandbrakeCLI but couldn't get the x265 encoder to choose anything other than AVX2. It seems to work only in the GUI which I can't automate in a benchmarking script.

 

[igor_kavinski]

A fascinating article

Definitely! Reading that makes me think that one of the reasons for disabling AVX-512 on Alder Lake may have been that they knew they were going to try for frequencies north of 5.5 GHz with Raptor Lake and AVX-512 throttling could have led to massive frequency dips of 1 GHz or more. I wish someone kind and wealthy (hard to find but they exist!) would donate more recent hardware to the Chips and Cheese guys like the W7-3545X. Reading about 10900X in 2025 is kinda sad.

 

[yuri69]

That C&C article is very in-depth but... it repeatedly compares against Intel's first gen AVX-512 - Skylake-X. That means a 2024 core vs a 2017 core which is rather unfortunate given there are Cooper Lake, Ice Lake, Sapphire Rapids, Emerald Rapids, and Granite Rapids.

 

[MS_AT]

Definitely! Reading that makes me think that one of the reasons for disabling AVX-512 on Alder Lake may have been that they knew they were going to try for frequencies north of 5.5 GHz with Raptor Lake and AVX-512 throttling could have led to massive frequency dips of 1 GHz or more. I wish someone kind and wealthy (hard to find but they exist!) would donate more recent hardware to the Chips and Cheese guys like the W7-3545X. Reading about 10900X in 2025 is kinda sad.

Intel solved Skylake problems at least partially with IceLake. I would expect that Raptor Lake would do no worse than Zen5. https://travisdowns.github.io/blog/2020/08/19/icl-avx512-freq.html#rocket-lake

 

[Win2012R2]

[deleted since that was already mentioned]

 

[Win2012R2]

I would expect that Raptor Lake would do no worse than Zen5

Xeons had double FMA units and clients were cut to 1, where as Zen 5 is full featured - that could make a big difference to code that uses those.

 

[MS_AT]

Xeons had double FMA units and clients were cut to 1, where as Zen 5 is full featured - that could make a big difference to code that uses those.

I meant when it comes to maintaining the frequency. Not performance The biggest problem of Skylake implementation was that it was not worth it to sprinkle AVX512 into the code, now it is not a problem with Zen4/5 and should not be with newer Xeons since client CPUs from Intel lack avx512.

And about Xeons:

Across all of the benchmarks run with AVX-512 on/off, when AVX-512 was enabled for the Xeon Platinum 8380 2P "Ice Lake" server and running these AVX-512 heavy workloads it saw a ~175MHz drop on average to the peak CPU frequency. Now with Sapphire Rapids this is no longer the case but the peak CPU frequency was similar with/without AVX-512. As shown previously, AMD EPYC Genoa also doesn't experience the AVX-512 downclocking. With Ice Lake it wasn't too bad to begin with unlike the AVX-512 Skylake days.

after https://www.phoronix.com/review/intel-sapphirerapids-avx512/8, of course it's not the level that C&C go into, but it supports the case that Intel has improved the behaviour since Skylake-X.

 

[Win2012R2]

The biggest problem of Skylake implementation was that it was not worth it to sprinkle AVX512 into the code

Yep, it was crazy - only ok for sustained workloads.

 

[OneEng2]

Ironic that Intel brought these instructions to x86 in an effort to find an advantage over AMD and it has ended up being an advantage FOR AMD.

Nice set of articles. Super informative. Thanks!

 

[Jan Olšan]

Unfortunately, it doesn't work for x265 which is what people are more interested in. I spent considerable time trying different variations of that command parameter with HandbrakeCLI but couldn't get the x265 encoder to choose anything other than AVX2. It seems to work only in the GUI which I can't automate in a benchmarking script.

It absolutely should work with x265 commandline but you are not using that.
If you are using a different frontend, be it ffmpeg or handbrake cli, you need to look up the cli option to pass the encoder extra parameters.
Based on this I think you need --encopts asm=avx512

If the parameter is already in your commandline, add the ASM thing to the end, separated by colons

-x, --encopts <string>  Specify advanced encoding options in the same
                           style as mencoder (all encoders except theora):
                           option1=value1:option2=value2

Also, Handbrake doesn't let you copy the resulting commandline of the encoding job (or show it in the logs)?
I fondly remember MeGUI for that. It casually taught you to really use the commandline encoder without fluff, by basically constructing the command for you (and having the various encoder options really well commented...). It also worked by calling the original encoder executables, not its own shim binary, so problems like this were avoided.

 

[igor_kavinski]

It's funny that AVX-512's need arose due to a software renderer: https://www.gamedeveloper.com/game-platforms/rad-launches-pixomatic----new-software-renderer




The 9950X's full width AVX-512 implementation is almost the 20 year anniversary of Pixomatic!

One wonders if Intel hadn't messed up AVX-512's rollout to consumer CPUs, we could have had a modern software renderer for games written in AVX-512, giving us the chance to enjoy Cyberpunk 2077 with Minecraft like pixelated graphics

Maybe some day average Joe's 128 thread consumer CPU WILL run games in software mode at acceptable speed and graphics quality and we won't have to bother with a mandatory fat GPU anymore.

 

[lamedude]

Nearly five years ago Mike Sartain and I had just put the wraps on our x86 software renderer, Pixomatic. We had done everything we could think of to
speed it up, and while it had certainly gotten a lot faster, it was still so much slower than hardware that we knew we could never close the gap. As we were
setting up in the RAD Game Tools booth at Game Developers Conference one morning, I said to Mike: "Man, if only Intel had a lerp [linear
interpolation] instruction!"

Mike pointed across the aisle at the Intel booth. "Maybe you should ask for one."

The odds seemed long, to say the least, but I didn't have any better ideas, so I went over and talked with Dean Macri, our developer rep. That resulted in a
couple of maverick Intel architects, Doug Carmean and Eric Sprangle, coming over to chat with us later; and somehow, over the course of five years, that
simple question led to a team at RAD -- which grew to include Tom Forsyth and Atman Binstock -- working with Intel to help design an instruction set
extension and write a software graphics pipeline for it.

https://www.cs.cmu.edu/afs/cs/academic/class/15869-f11/www/readings/abrash09_lrbni.pdf