We found the Missing Performance: Zen 5 Tested with SMT Disabled

[Steevo]

So it’s really either AMD needs to step up and release software to control thread placement or update their microcode or BIOS and work with MS to get it fixed

 

[evernessince]

Significant, when there's a "mere" 2.5% average gain between 7700X and 9700"X", otherwise. So this small tweak effectively doubles the performance difference, and brings gaming much closer to expectations.

That's more a testament to how little the 9700X improved over the 7700X. Double 2.5% is still only 5%. At the cost of lower overall application performance to boot. Anyone chasing gaming FPS are just going to get a 7800X3D. It'd be one thing if we were talking class leading gaming performance but we aren't.

It's also only slower in some applications.

Perfect example of a Truism. It's correct but adds nothing to the conversation. We are all aware that averages are comprised of mixed results but that doesn't change the fact that overall disabling SMT reduces performance in non-gaming scenarios. You are aware of this because you used an average in your comment, me interjecting that in fact part of that 2.5% figure includes benchmarks when Zen 5 regresses in performance is worthless in this context as your comment is here.

 

[kiniku]

On the Ryzen 9 7950X3D, AMD put the 3DV Cache on one CCD only, so a software driver hooked into Windows Game Mode detection to detect games and push those onto the big-cache-cores, while applications end up on the high-frequency cores. The problem is that this solution is quite brittle and that it has a lot of moving parts that can and will break, especially the ones designed by Microsoft.

I never saw it this way. Less moving parts, fewer points of potential failure. And this AMD/Microsoft dual software method of exploiting 3DV cache? I'll pass, next time. I think I may go to Intel if they improve the efficiency of their next gen CPUs. Furthermore, I couldn't care less about 20 more FPS in my gaming. Since I game at 3440 and above, I'll compensate with GPU power.

 

[phanbuey]

So it’s really either AMD needs to step up and release software to control thread placement or update their microcode or BIOS and work with MS to get it fixed

New architecture and AMD always takes a few months to get going. Reveiws on beta bios were barely stable.

 

[RandallFlagg]

<snip>

In Intel Emeral Rapids I can get 64 core per CPU or I can get 128 Core from AMD in a single socket. So in a 2U design it will be common to see 128 Intel vs 256 AMD cores. Now lets buld out a data center on their scale. How many Us are now required for say 15k cores.

<snip>

Did you bother to do the math on that?


A 7U Cisco UCS can handle 7 x410c compute nodes with 4 sockets of 60 core Xeons, which comes out to 1,680 cores.

So you are talking about 70U to get to 16,800 cores, or about 2 racks.

Meanwhile, If you have 20% more performance per core, you can take those 15,000 cores and reduce them to 12,000, and with something like SQL Server you are talking about $36,000,000 difference in licensing costs.

Hopefully you are working for the competition (3000 cores X $12,000 per core = $36,000,000).

Software cost virtually *always* dwarfs hardware costs.

 

[EmmanuelMar]

Games are very relative, some are poorly designed to work with a certain cpu/gpu and you achieve more performance just because of the brand of cpu/gpu, not to mention that these games are limited to use a certain number of threads, this can be change in some games so that instead of using 4 threads they use 6 or 8 threads and naturally the performance increases, the important thing is the synthetic tests and I would not really disable SMT because you will lower the multithread performance.

 

[Gooigi's Ex]

Excellent article and now has created a question. What would be the performance of Windows handhelds when you disable SMT? Window Handheld PCs don’t play at 4k at all and barely into 1440p and feels nice at home at 900/1080p gaming.

 

[Panther_Seraphin]

Did you bother to do the math on that?


A 7U Cisco UCS can handle 7 x410c compute nodes with 4 sockets of 60 core Xeons, which comes out to 1,680 cores.

So you are talking about 70U to get to 16,800 cores, or about 2 racks.

Meanwhile, If you have 20% more performance per core, you can take those 15,000 cores and reduce them to 12,000, and with something like SQL Server you are talking about $36,000,000 difference in licensing costs.

The one "weakness" of AMD at the moment is the lack of Quad socket support, but the physical size of the processors would realistically relegate them to extreme niche cases.
A dual 9684X system I bet would put that quad socket through its paces however. The only thing it would really lack is the gargantuan amount of RAM that you could stuff in one of those blades..............64 Dimms vs 24 for the AMD.

*Snip*

Software cost virtually *always* dwarfs hardware costs.

*Snip*

Definately dont disagree there, especially when things are moving towards a per thread let alone core, per year sort of landscape moving forward. Some of the quotes I have seen for things have made me physically wince and go HOW MUCH?

 

[trparky]

So, maybe AMD should come out with a chip that's like the 9700X with one chiplet where it's an eight-core chiplet that has no SMT support and another chiplet that a whole bunch of Zen 5c cores to make up for the loss of SMT.

 

[Suspecto]

You didn't find the missing performance, you found meaningless 2% which will probably be an issue with the Windows scheduler and Agesa.

 

[trparky]

You didn't find the missing performance, you found meaningless 2%.

I'd have to agree with you on that. The low power profile of this chip is what's holding it back. Like I mentioned in the other 9700X thread, AMD really needs to include a "Take the gloves off" mode.

 

[Vya Domus]

I know some people will say that core density is all matter but that is not true at all as many server software use a per core license.

That is huge and largely ignored by the desktop crowd.

In many cases it's simply unavoidable, you have to pay the licensing fee per core anyway and if you have to then you might as well pick the most cost effective platform which will typically be the one that will give you the most cores per node, core density is still a major factor in almost every datacenter application. You're assuming that 10% faster cores allows a reduction in core count by 10% but the scalability simply doesn't work like that a lot of the times, if ever.

But even if it did, core density would still matter. Take your example, databases do not scale well with cores, this isn't anything new but there's major problem with this "least amount of cores possible approach". Because of the way queries work a lot of the time they will block the execution of subsequences queries, so the longer each query takes on average the worse the performance will get. If you're trying to cram as many connections per core then you're likely trading speed for whatever you saved up for your licensing costs. The question is how much cost can you cut until you can no longer tolerate the performance hit, it's definitely not 10%, that's for sure. The multithreading in databases works locally for each instance, so performance wise you're still better served by more cores/node.

 

[trparky]

I'd have to agree with you on that. The low power profile of this chip is what's holding it back. Like I mentioned in the other 9700X thread, AMD really needs to include a "Take the gloves off" mode.

And it seems that I'm not the only one that thinks this, der8auer thinks that way as well.

Very Efficient Ryzen 7 9700X Held Back by Power Limits!

 

[W1zzard]

The low power profile of this chip is what's holding it back

Red bar has power limits removed, plus PBO, plus Curve Optimizer

 

[Braegnok]

Ryzen 9700X running new CKD memory straps is looking solid. https://www.overclock.net/posts/29354757/

After the 15th we will see 9950X results,.. the 9700X running 8800/2200 synced profile is amazing.