AMD's Compressonator 4.4 Adds AVX-512 Support

[phoronix]

Phoronix: AMD's Compressonator 4.4 Adds AVX-512 Support

The AMD Compressonator open-source tool suite that is under the GPUOpen umbrella has now added AVX-512 support alongside other enhancements in its v4.4 update...

https://www.phoronix.com/news/AMD-Compressonator-AVX-512

[rene]

cool, can't wait to test on the Via CHA-001 prototype w/ AVX-512 https://www.youtube.com/watch?v=1PSmbjKsXuE

[tildearrow]

Typo:

GUI drivem

[ms178]

I am still not sure if AVX-512 or more cores are of more value for day-to-day usage, e.g. compilation or general desktop use. Come on, Intel and AMD - can't you offer higher core counts and AVX-512 support together in a single consumer product?! Having to chose between these features makes current offerings of you uninteresting.

[Anux]

can't you offer higher core counts and AVX-512 support together

? Every Zen 4 independent of it's core count has AVX512.

[coder]

I am still not sure if AVX-512 or more cores are of more value for day-to-day usage, e.g. compilation or general desktop use.

Well, in day-to-day use, AVX-512 is virtually irrelevant. So, if you're trying to determine the cost/benefit ratio, then we'd have to conclude it's a waste of die space for most people. However, the main reason it's there isn't for the typical user. Rather, it's for the cloud and hyperscalers. If having AVX-512 on there gives AMD more volume, then they can potentially reduce costs enough that it's actually "free" to you.

Looking at the abstract question of more cores vs. AVX-512, you'd have to account for how much space it's adding. I didn't find an answer to that question, but according to what we know, AMD was more careful to implement it in an area-efficient manner than Intel. This analyzes their implementation further, in case you're curious:

https://chipsandcheese.com/2022/11/0...cution-engine/

If you compare that to Intel's gen 12 and 13 CPUs, they have the worst of both worlds - AVX-512 is chewing up space, but you get literally zero benefit from it, because it's completely disabled!

Come on, Intel and AMD - can't you offer higher core counts and AVX-512 support together in a single consumer product?! Having to chose between these features makes current offerings of you uninteresting.

Why do you need more than 32 threads? If you're really torn over the issue of AVX-512, then I'd say just go with AMD. The 7950X is close enough to the i9-13900K that you're not giving up much.

[ms178]

Why do you need more than 32 threads? If you're really torn over the issue of AVX-512, then I'd say just go with AMD. The 7950X is close enough to the i9-13900K that you're not giving up much.

I come from a Haswell-EP 18-Core with 3.8 Ghz Turbo and quad-channel DDR3-RAM. This is still a capable chip for multi-threaded workloads and I got these components for a great bargain but the idle power consumption is literally burning money nowadays as power is very expensive where I live. I need the multi-threaded performance for code compilation mostly as I am frequently compiling parts of the Linux stack for my system. That aside, gaming is also a priority.

AVX-512 or the X3D parts would be interesting for a x86-64-v4/Zen 4 build of CachyOS that is coming later this summer. But as I don't do HPC/AI workloads, I'd say I could live without these two features for my next build that should last me at least for the next four years. I also still have 32GB DDR4-3000 (Micron Rev. E - so it can be OCed to 3600 GT/s with ease) around that I'd like to use (from a failed AM4 project that was my badest PC platform experience ever, hence the risk aversiveness to jump on AM5).

Zen 4 lacks the multi-thread price competitiveness in the mid-tier, is DDR5-only, stability is still in a Beta state and pricing is also not great for the platform. And I don't give too much on upgrade promises from AMD either after the community had to force them to change course on X370/B450 support (and history with AM3/AM3+ in the past).

14th Gen seems to tick more of my boxes. It is a bug-fix revision on a mature platform where I can still use my DDR4-RAM, it brings some decent core count increases in the i5 and i7 segment, which would be a worthy upgrade for my ageing Haswell. The only downside I can think of are power consumption/high temps and pricing. But hopefully both of these are less of a concern in the i5 - i7 segment and just ordered a decent Z690 DDR4 board today that brings PCIe 5.0 at 149 EUR.

[coder]

I come from a Haswell-EP 18-Core with 3.8 Ghz Turbo and quad-channel DDR3-RAM.

What speed? DDR3-1866? That'd give you nominal throughput of 59.7 GB/s.

Compare to dual-channel DDR5-5200 (Ryzen 7000's official limit) corresponds to 83.2 GB/s. Or, a mild overclock to DDR5-5600 (the fastest ECC UDIMMs I've seen) gives you 89.6 GB/s. So, you don't need quad-channel for that kind of bandwidth. Plus, DDR5 has now bifurcated the channel width, giving you quad-channel parallelism across 2x 64-bit DIMMs.

I got these components for a great bargain

Now is a good time to buy!

but the idle power consumption is literally burning money nowadays as power is very expensive where I live.

You can modify the default power limits in the BIOS. When you do this, Zen 4 loses very little performance. This shows an all-thread workload, where the 125 W setting delivered 96.0% of the 230 W default. And 105 W gives you 93.6% of the performance:

Source: https://www.anandtech.com/show/17641...13900k-7950x/2

I also still have 32GB DDR4-3000 (Micron Rev. E - so it can be OCed to 3600 GT/s with ease) around that I'd like to use

Well, Anandtech found that DDR4 can considerably hurt your multithreaded performance on Alder Lake. Raptor Lake has a bit more cache, but then it's also got more cores to feed...

Source: https://www.anandtech.com/show/17047...d-complexity/8

You can also just do the math. DDR4-3000 -> 48.0 GB/s. DDR4-3600 -> 57.6 GB/s. That's less bandwidth than you currently have in your Haswell-E box.

14th Gen seems to tick more of my boxes.

It's more power-hungry. Yes, you can tame it by restricting power limits, but you give up more performance than with Zen 4.

[Anux]

but the idle power consumption is literally burning money nowadays as power is very expensive where I live.

You should know that idle power is not better with current PCIe5/DDR5 high end plattforms (around 80W). With a Threadripper you will easily be over 100W idle. Ofcourse the board has a big impact and can change idle of the same components +-20W.

from a failed AM4 project that was my badest PC platform experience ever, hence the risk aversiveness to jump on AM5).

Then you should either sell DDR4 or build a plattform with it (AM4 is really cheap now and everyone sells their used AM4 for even cheaper). Keep in mind what coder said about DDR4.
With a high end board and 5950X you might still sit at 60W idle. Intel with DDR4 might be better at idle. If you start overclocking RAM, expect your idle to go up significantly, stay at 1.2 V for best idle.

And I don't give too much on upgrade promises from AMD either after the community had to force them to change course on X370/B450 support (and history with AM3/AM3+ in the past).

You have to be a little picky with your motherboard manufracturer but AMD always delivered on their promise, I have upgraded and recycled lots of boards/CPUs over the decades. 1 or 2 CPU upgrades per plattform and many bucks saved.

You can modify the default power limits in the BIOS.

But that doesn't change idle power.

It's more power-hungry. Yes, you can tame it by restricting power limits, but you give up more performance than with Zen 4.

There are no benchmarks for 14th gen yet? Yes its just a 13th gen update but maybe efficency has improved. But money wise it's not a good idea to buy the newest CPU after release, you will hardly make up for it with your power bill compared to a 13th gen or AM4.

[coder]

There are no benchmarks for 14th gen yet? Yes its just a 13th gen update but maybe efficiency has improved.

Same process node and no microarchitecture changes. So, I'm not expecting any efficiency gains (unless they add yet more E-cores).

I could be wrong, but I think they'll probably just juice clock speeds and maybe cache sizes enough for the top SKU to catch the 7800X3D in most games. If anything, that'll hurt efficiency.