Info 64MB V-Cache on 5XXX Zen3 Average +15% in Games

[Kedas]

Well we know now how they will bridge the long wait to Zen4 on AM5 Q4 2022.
Production start for V-cache is end this year so too early for Zen4 so this is certainly coming to AM4.
+15% Lisa said is "like an entire architectural generation"

 

[Joe NYC]

They basically announced 3 new server CPUs over the next 18 months. Crazy when you consider how competitive Milan already is. Sapphire Rapids needs to be really good for Intel.

Hopefully the desktop gets more than the scraps. But with the repeated delay or disappearance of Chagall it seems they are focused on servers.

AMD made hard right turn since May announcement of V-Cache, when AMD said Ryzen would be the launch platform of V-Cache, only to put desktop Ryzen on a side burner and allocate all the dies to Milan X.

I am guessing that the negotiations between AMD and Microsoft Azure were especially fruitful, and justified abandoning desktop Zen3D for now...

Windows 11 fiasco also posed additional turbulence, something best to avoid and change plans...

 

[Joe NYC]

I've seen it referenced a few times on these forums that there is a SLIGHT latency hit from 3d stacking the L3 cache. It's either related to having expanded tables for the larger size, or the fact that the electrical distance from each core to each L3 cell is still physically slightly longer (though, we're still talking extremely tiny amounts. With the clock speeds associated, it does matter).

The explanation I have seen is that on Milan X, latency for the base 32 MB is unchanged. But when accessing V-Cache, there is some (~10% ?) latency hit.

But the way to look at it is not to compare the 64 MB access with 32 MB access, but 64 MB V-Cache access with main memory access.

 

[Joe NYC]

Yeah, I'm sure in edge cases there will be a huge jump in performance, way higher than that 15% mark.

I'm just wondering, is there a point of diminishing returns with huge caches? Or is it literally the more the merrier?

I'm aware that latency has a huge role to play in this, but I'm asking in the assumption that latency differences are minimal, as it seems to be from the latest leaks.

If the application is running nearly completely out of cache, there is nothing more to gain.

But if the application has memory accesses and cache misses, there will be gains.

The rule of thumb used to be that every doubling of the size of the cache gave you approx. same increment of performance. Which means that one side has to rise exponentially for the other side to rise linearly. But that is just a general guide, each application will behave differently...

 

[epsilon84]

If the application is running nearly completely out of cache, there is nothing more to gain.

But if the application has memory accesses and cache misses, there will be gains.

The rule of thumb used to be that every doubling of the size of the cache gave you approx. same increment of performance. Which means that one side has to rise exponentially for the other side to rise linearly. But that is just a general guide, each application will behave differently...

Yeah, that makes sense. I know this is looking at 10th gen Intel CPUs, but its quite interesting to see the variance in gaming performance between the different cache sizes, everything else being equal:

Some games like BF:V actually gain close to 20% from 12MB to 20MB L3, but others (maybe due to a GPU bottleneck) show almost no gains. I have a feeling Zen 3D will exhibit similar traits, with some huge uplifts in certain games followed by others with minimal gains. Hence that 15% 'average' figure.

 

[Joe NYC]

Yeah, that makes sense. I know this is looking at 10th gen Intel CPUs, but its quite interesting to see the variance in gaming performance between the different cache sizes, everything else being equal:

Some games like BF:V actually gain close to 20% from 12MB to 20MB L3, but others (maybe due to a GPU bottleneck) show almost no gains. I have a feeling Zen 3D will exhibit similar traits, with some huge uplifts in certain games followed by others with minimal gains. Hence that 15% 'average' figure.

I am currently playing a game (whit a special Mod) that I think would gain quite disproportional uplift from Zen3D. It is a strategy game that is repetedly traversing gigabytes of memory.

(It's out inopen alpha if anyone is interested)
Steam Workshop::MEIOU and Taxes (steamcommunity.com)

 

[deasd]

I couldn't find a reliable source which I saw before that compare APU & CPU IPC, on the basis of their different L3 size...
I'm not sure how L3 could affect ST perf but I believe there's a significant bottleneck in L3 when a stress MT workload kicks in.
In GB5 there's a huge difference between ST&MT(11% vs 20%) and it's very likely not only generated by the clock difference(4.4Ghz vs 4.6Ghz).
So I guess a MT workload or even when triggering any multicore processing the L3 size would become a very sensitive factor.

 

[DrMrLordX]

I think that even after a no show of Zen 3D when it mattered,

Zen3D isn't late.

 

[Joe NYC]

Zen3D isn't late.

I wasn't clear, Zen 3D is fine ahead of schedule, running on Azure Cloud.

But desktop Ryzen 3D missed the ideal window to launch, from which it will not recover, PR-wise. It will take a lot more to regain the performance crown than what would have taken to keep it, PR-wise.

 

[therealmongo]

Opinion does not equal a fact !

 

[Mopetar]

I wasn't clear, Zen 3D is fine ahead of schedule, running on Azure Cloud.

But desktop Ryzen 3D missed the ideal window to launch, from which it will not recover, PR-wise. It will take a lot more to regain the performance crown than what would have taken to keep it, PR-wise.

What are you on about? When Zen 3D comes out it will be close enough to Intel in most regards that if you're already on AMD's platform and looking to upgrade from Zen 2 without dropping a lot of money for a new board and DDR5 you've got a solid option.

Zen 4 shouldn't have any trouble taking a notable lead over Intel's offerings at the moment. How many people are going to jump on Zen 3D as complete newcomers when there was always the option of waiting for Zen 4, especially if you're building a new system as opposed to upgrading.

I think you have some fundamental misunderstanding of the purpose of Zen 3D as a product and are drawing bad conclusions from bad assumptions.

 

[Ajay]

Zen4 is 25% higher performance per Core, you add more Cores and higher bandwidth you can still see Zen4 EPYC>>MilanX>Milan for general purpose CPUs. For sure there will be Zen4 3D Cache options down the line for special purpose CPUs

You keep saying this when it hasn't been confirmed! Please link a source or stop posting this.

 

[nicalandia]

You keep saying this when it hasn't been confirmed! Please link a source or stop posting this.

 

[biostud]

View attachment 52703

But do we know if zen4 comes without 3d cache from start?

 

[nicalandia]

But do we know if zen4 comes without 3d cache from start?

No, Milan X is a special CPU, there will be Genoa-X just before Bergamo is released.

 

[maddie]

View attachment 52703

That's the process not the products. Read the fine print at bottom. It even has a hedge with "goal improvements are".

 

[nicalandia]

That's the process not the products. Read the fine print at bottom. It even has a hedge with "goal improvements are".

Yeah, I see that thanks for pointing that out, it should be>1.25 so that could be the least at ISO power

 

[Saylick]

25% performance seems to align with this figure for N5... It appears N5 eLVT has higher standby power than N7 uLVT to accomplish that 25% increase, while N5 HPC uLVT is roughly 25% at iso-power. There have been rumors of TSMC developing a tailored process for AMD, so this could be the fruits of that labor, i.e. N5 HPC.

 

[moinmoin]

25% performance seems to align with this figure for N5 HPC

For N5 eLVT you mean. N5 HPC eLVT would add another 10% on top of that according to your quoted graph (thanks!).

 

[DrMrLordX]

But desktop Ryzen 3D missed the ideal window to launch

Desktop Zen3D isn't late.

 

[Ajay]

View attachment 52703

Well, as I said before, it doesn't say per core. Since this is HPC, we don't know what this vague '1.25x perf' corresponds to. I hope Zen 4 gives 25% performance per core, or more. It's just not what AMD is saying right now.

 

[Joe NYC]

What are you on about? When Zen 3D comes out it will be close enough to Intel in most regards that if you're already on AMD's platform and looking to upgrade from Zen 2 without dropping a lot of money for a new board and DDR5 you've got a solid option.

Zen 4 shouldn't have any trouble taking a notable lead over Intel's offerings at the moment. How many people are going to jump on Zen 3D as complete newcomers when there was always the option of waiting for Zen 4, especially if you're building a new system as opposed to upgrading.

I think you have some fundamental misunderstanding of the purpose of Zen 3D as a product and are drawing bad conclusions from bad assumptions.

Intel Rocket Lake CPUs sales are being lifted, selling ahead of Zen CPUs in the last week.

Did Rocket Lake suddenly became materially better, more competitive vs. Zen?

No, Rocket Lake sales are up vs. AMD CPUs because AMD did not have Zen 3D in comparisons against Alder Lake.

Maybe it does not make sense to you. Maybe you should think a little harder, until it clicks.

 

[Joe NYC]

But do we know if zen4 comes without 3d cache from start?

We don't know, but I am going to guess no, not from the start. Not in volume, maybe some low volume test runs, at best.

 

[nicalandia]

No, Rocket Lake sales are up vs. AMD CPUs because AMD did not have Zen 3D in comparisons against Alder Lake.

Amd could care less about Rocket lake sales going up when compared to Desktop Ryzen, Their priority is the Client market and they just had a few big wins recently, they can charge premium per CCD as compared to desktop where people made a fuss about a $50 increase for for the 5000 line


Meta AMD North Dome 36 Core 95W TDP CPU

Facebook Meta AMD EPYC North Dome CPU and Platform Details

Details on the big AMD EPYC win at Facebook-Meta were shown with the 36C 95W TDP EPYC in the North Dome OCP platform

www.servethehome.com

 

[moinmoin]

Facebook Meta AMD EPYC North Dome CPU and Platform Details

Details on the big AMD EPYC win at Facebook-Meta were shown with the 36C 95W TDP EPYC in the North Dome OCP platform

www.servethehome.com

That has to be a rather odd custom configuration. Only way I see an Epyc chip reach 36 cores is with 6 chiplets with 6 active cores each. Wonder what about that makes it more suitable to 95W TDP than other possible configurations.

 

[UsandThem]

Never mind.....not worth it.