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"

 

[biostud]

I found it... Yes, saying AMD is behind Intel in every respect is delusional. As I said before, (to surmise)

Server: AMD way out front
HEDT: way out front
Desktop: contested, depends on the application, but Intel has a power problem.
Deskptop: This is contested, and AMD in in the fight, with new stuff coming shortly.

Your response ???

My response: By Deskptop, do you mean Laptop?

 

[Markfw]

My response: By Deskptop, do you mean Laptop?

Yes, typo.

 

[nicalandia]

Yes, typo.

Your point still stand.

Intel is Delusional(the CEO No less) thinking that AMD will forever be at the "Rearview Mirror". They were bullied out of the UHEDT by ThreadRipper. Their latest 10nm Xeon can't compete with ThreadRipper Pro(Zen2 based) All the while Milan based EPYC are baby seal clubbing(Never seen a one side beat down so bad) and their only hope to stop such beating which is sapphire rapids Xeons have been delayed while Milan-X is showcasing their 3D $ Prowess. Zen4 will only make things worse for many lines of Intel by the end of the year

 

[nicalandia]

Updated Info from Chips and Cheese. Impressive performance...


Deep Diving Zen 3 V-Cache

Deep Diving Zen 3 V-Cache

This is the deeper dive of AMD’s V-Cache that we teased with our short latency article and we will be covering a little more on the latency front along with the bandwidth behavior of V-Cache …

chipsandcheese.com

 

[Saylick]

Updated Info from Chips and Cheese. Impressive performance...

Thanks for linking. Looks like it has the answers to the cache bandwidth question that you were looking for earlier.

 

[yuri69]

Updated Info from Chips and Cheese. Impressive performance...

The top perf gain among their benchmarks was 7.6% compared to standard Milan. You simply need a workload profiting from the large cache.

 

[DrMrLordX]

Updated Info from Chips and Cheese. Impressive performance...

Considering all they did was strap on some extra L3, yeah, I'd say they did pretty well. If that had been an attempt at a generational leap then not so much.

 

[JoeRambo]

1-CCD performance increase is rather disappointing in absolute numbers. Some of those tests are memory subsystem darlings, like compression, compile, heavy FP that are supposed to scale real well with L3 size/BW.

Still 10% increase clock for clock for the effort spent in the same generation is good result. I think what is happening is CCD choking on memory link to IOD, not enough bandwidth to prefetch to L3 and there is not much use of increased capacity even in tests like compression. Would get fixed by DDR5 and widened bw between CCD/IOD that is happening in Zen4.

What these benchmarks don't show, is so called "socket/system" wide performance. It is hard to quantify without running things like SPEC or SPEC java benchmarks. Or even better - having both Z3 and Z3X instances and having detailed performance metrics for your services.
Esp cloud vendors will love these chips, as larger L3 can shield the rest of the system from busy JVM that is for example running garbage collection and would hurt memory latency/bw for the rest of chip or even system by loading intersocket links to reach memory connected to different socket. "Loaded" memory latency is rather different from unloaded and single busy CCD can have knock on effect on whole system performance.

 

[nicalandia]

Considering all they did was strap on some extra L3, yeah, I'd say they did pretty well. If that had been an attempt at a generational leap then not so much.

It's just extra Cache for apps that need it and 15% improvement is no longer considered a generational leap any more? Oh how spoiled have AMD Made a bunch of you guys. Remember the 3-5% IPC increase Intel was feeding us?

 

[nicalandia]

1-CCD performance increase is rather disappointing in absolute numbers. Some of those tests are memory subsystem darlings, like compression, compile, heavy FP that are supposed to scale real well with L3 size/BW.

Still 10% increase clock for clock for the effort spent in the same generation is good result. .

That is just the tests shown by Chips and cheese. The one shown by Microsoft the Performance can be Quite exceptional like 78% performance increase(50% average)

 

[DrMrLordX]

It's just extra Cache for apps that need it and 15% improvement is no longer considered a generational leap any more?

I saw about 7%. Where are you seeing an average of 15%?

 

[nicalandia]

I saw about 7%. Where are you seeing an average of 15%?

You need to get your eyes check then.

15% Average IPC in gaming
50% Average IPC in HPC
And 12.5% IPC on C&C

 

[DrMrLordX]

You need to get your eyes check then.

Page 71 - Info - 64MB V-Cache on 5XXX Zen3 Average +15% in Games

Page 71 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

The top perf gain among their benchmarks was 7.6% compared to standard Milan. You simply need a workload profiting from the large cache.

@nicalandia , you were saying?

 

[biostud]

That is just the tests shown by Chips and cheese. The one shown by Microsoft the Performance can be Quite exceptional like 78% performance increase(50% average)

View attachment 56340

No wonder Microsoft is scooping up these processors. That is quite an improvement, in performance/watt and raw performance.

 

[nicalandia]

@nicalandia , you were saying?

Your grasping at straws dude.

 

[DrMrLordX]

Your grasping at straws dude.

How??? It's right there in the Chips & Cheese benchmark.

OpenSSL was a regression for Milan-X

Gem5: +7.6%
libx264: +4.6%
7-zip compress: +4.9%
y-cruncher (2.5b): +1.4%

These numbers have to be taken in some context, since apparently the Milan numbers used for comparison were from a cloud instance, so clocks might not have varied with the current draw of the workload in the same way as the Milan-X sample. But it's the only actual data we have from an independent reviewer. I'm hoping they can get some bare metal Milan numbers for comparison and also run a few more benches, such as 3d-rendering? I do not expect them to produce gaming results but you never know.

 

[nicalandia]

How??? It's right there in the Chips & Cheese benchmark.

Not at ISO Speed. VM vs Baremetal. At ISO speed the IPC is 12.5% on those instances, 50% on Microsoft Testing and 15% in Gaming. How is that not impressive or even Not "Generational" ?

 

[Ajay]

How??? It's right there in the Chips & Cheese benchmark.

OpenSSL was a regression for Milan-X

Gem5: +7.6%
libx264: +4.6%
7-zip compress: +4.9%
y-cruncher (2.5b): +1.4%

These numbers have to be taken in some context, since apparently the Milan numbers used for comparison were from a cloud instance, so clocks might not have varied with the current draw of the workload in the same way as the Milan-X sample. But it's the only actual data we have from an independent reviewer. I'm hoping they can get some bare metal Milan numbers for comparison and also run a few more benches, such as 3d-rendering? I do not expect them to produce gaming results but you never know.

What are the targeted workloads for Milan-X? I suspect the benchmarks you list aren't pertinent to the market Milan-X is selling into.

 

[Kedas]

The 15% is the average of 5 games shown in AMD's slide.
We all know the advantage will be very load dependant, even in games we see a small change for League of Legends but a big +25% for Monster Hunter World both DX11.

For the relevant changes on the Desktop software spectrum we will have to wait for a full review.

 

[DrMrLordX]

Not at ISO Speed.

What do you expect? The 5800X3d also won't be "at iso speed" versus the 5800X, as has already been advertised. Also, in case you weren't paying attention, I already mentioned that the Milan instance may have had elevated clocks during at least one of the benchmarks (y-cruncher).

What are the targeted workloads for Milan-X? I suspect the benchmarks you list aren't pertinent to the market Milan-X is selling into.

It's the only Zen3d anyone has tested independently. Nobody has their hands on a 5800x3d. This data is the closest we can get right now to knowing how a Zen3d part might improve performance for us in desktop applications. A lot more testing needs to be done, and we need better samples to get a clearer picture. But it's still a darn sight better than relying on some vendor benchmarks/advertising.

For the relevant changes on the Desktop software spectrum we will have to wait for a full review.

Generally agreed.

 

[nicalandia]

The 15% is the average of 5 games shown in AMD's slide.

10 Games, two slides. Same 15% increase.


So far these are the lists of games with 15%+ performance improvements(3D-V Cache) over a stock 5900X and 12900KS Alder Lake. Pay attention to the Tie/Even games where it does the same with the 12900K so it must be a frame cap or something. Also CS:GO Three of the CPUS ties, the 5900X, 12900K and 5800X3D so the issue is on the game.

DOTA 2: 18% Over 5900X - No info on 12900K
Monster Hunter World: 25% over 5900X - No info on 12900K
League of Legends: 4% Over 5900X - No info on 12900K
Fortnite: 17% Over 5900X - No info on 12900K
Final Fantasy XIV: 20% Over 5900X - 20% Over 12900K
Shadow of the Tomb Raider: 10% Over 5900X - 10% Over 12900K
Far Cry 6: 20% Over 5900X - 11% Over 12900K
Gears V: 12% Over 5900X - Tie/Even with 12900K
Watch Dogs Legion: 40% Over 5900X - Ties/Even with 12900K
CS:GO: Tie/Even with 5900X - Tie/Even with 12900K

 

[nicalandia]

What do you expect? The 5800X3d also won't be "at iso speed" versus the 5800X, as has already been advertised. Also, in case you weren't paying attention, I already mentioned that the Milan instance may have had elevated clocks during at least one of the benchmarks (y-cruncher).

You are tying to paint a bad light on 3D V Cache. It's a game changer and a generational leap. To say otherwise is just not accurate

 

[darkswordsman17]

Your point still stand.

Intel is Delusional(the CEO No less) thinking that AMD will forever be at the "Rearview Mirror". They were bullied out of the UHEDT by ThreadRipper. Their latest 10nm Xeon can't compete with ThreadRipper Pro(Zen2 based) All the while Milan based EPYC are baby seal clubbing(Never seen a one side beat down so bad) and their only hope to stop such beating which is sapphire rapids Xeons have been delayed while Milan-X is showcasing their 3D $ Prowess. Zen4 will only make things worse for many lines of Intel by the end of the year

View attachment 56220
View attachment 56222

To be fair, Gelsinger said in client/consumer, meaning not Epyc or Threadripper.

I also personally have a hunch his comment had little to do with performance and is more a harbinger of the return of Intel's old business practices. I fully expect Intel is gonna offer at or maybe even negative cost deals to OEMs for bundling Intel CPU+GPU in order to push AMD back out and try to gain GPU market share. I think his remark is that now that Intel is back competitive, there's no reason that OEMs can argue against the massive supply deals that Intel can offer. Being competitive helps protect such deals from investigation (not that I think Intel is worried about that, they've seen how it takes years for any such thing to even begin, but I'd guess it would help keep the OEMs from getting spooked about such).

 

[DrMrLordX]

. It's a game changer and a generational leap. To say otherwise is just not accurate

No, I'm trying to bring people back down to Earth until we get more verifiable 3rd-party data. Take the tribalism crap elsewhere. That one Milan-X is the only sample in the wild that anyone is testing publicly. It's all we've got.

 

[Mopetar]

The top perf gain among their benchmarks was 7.6% compared to standard Milan. You simply need a workload profiting from the large cache.

Not just the extra space, but a workload that specifically benefits from increasing the number of sets.

The performance gains are just a result of still having data that would have otherwise been evicted still in the cache. Anything that is only using memory once or twice is unlikely to benefit. Large data sets that don't have predictable access patterns are what should do best since more of that data can hang around in the L3 cache and benefit from being more accessible in the future.