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"

 

[eek2121]

Ha, AMD regarding 3D vcache for RDNA2:

06:19PM EDT - Q: Why Infinity Cache vs Stacked V-Cache A: Not discussing, were only evaluating on-die cache

 

[Hitman928]

AMD has already made their point abundantly clear. They won’t go 4+. The demo had 72mb and prod skus will have 192 (probably across 2 chiplets)

Pretty sure the demo had 192 MB which is 1 stack hi on 2 chiplets.

At the risk of misrepresenting another's argument, I understand the belief to be that AMD is sandbagging when they say the SKUs have 192 MB of cache and that if after ADL launches, ADL beats Zen 3 by enough in gaming, AMD will throw however many stacks they need to take back the lead instead of the 1 stack SKUs they keep mentioning.

 

[Ajay]

Bold of you to assume it will be present at all on single CCD parts. Remember, it is more expensive to make these chips. Margins are higher on the 5900X and 5950x.

Uh, I didn’t even say what you assumed I said 🙄

 

[eek2121]

Pretty sure the demo had 192 MB which is 1 stack hi on 2 chiplets.

At the risk of misrepresenting another's argument, I understand the belief to be that AMD is sandbagging when they say the SKUs have 192 MB of cache and that if after ADL launches, ADL beats Zen 3 by enough in gaming, AMD will throw however many stacks they need to take back the lead instead of the 1 stack SKUs they keep mentioning.

Read the hot chips presentation slides and get back to us. What I stated was fact, not an educated guess nor wishful thinking. One thing we can absolutely certain of is factual information in AMD presentations, for better or for worse.

 

[Joe NYC]

Pretty sure the demo had 192 MB which is 1 stack hi on 2 chiplets.

At the risk of misrepresenting another's argument, I understand the belief to be that AMD is sandbagging when they say the SKUs have 192 MB of cache and that if after ADL launches, ADL beats Zen 3 by enough in gaming, AMD will throw however many stacks they need to take back the lead instead of the 1 stack SKUs they keep mentioning.

Yes. And that's a fair representation of my argument, thanks.

 

[Joe NYC]

Anyone that thinks Zen 3 will beat Alder Lake in anything needs to go to AT bench, pick a ST benchmark of the 11900k, and add 15-20% to it, and see which number (Zen 3 or the new number) is higher. Hint: Alder Lake will be in most scenarios.

I agree. The numbers will likely be all over the place, but the gaming SKU Intel is launching Alder Lake with, 12900k will be convincingly higher than 11900K and Zen 3.

This Geegbench score analyzed by VideoCardz looks strong.
Intel Core i7-12700 non-K Alder Lake CPU is almost as fast as Ryzen 7 5800X in leaked Geekbench score - VideoCardz.com

 

[Joe NYC]

AMD has already made their point abundantly clear. They won’t go 4+. The demo had 72mb and prod skus will have 192 (probably across 2 chiplets)

Really? I must have missed that part.

As secretive as AMD is about everything, to the point of describing last years parts at conferences, not upcoming ones, why do you think AMD spilled everything about their V-Cache at the demo, left nothing for the launch?

 

[Hitman928]

Read the hot chips presentation slides and get back to us. What I stated was fact, not an educated guess nor wishful thinking. One thing we can absolutely certain of is factual information in AMD presentations, for better or for worse.

This is the slide I saw from their Hotchips presentation. Is there a different one I'm supposed to be looking for?

 

[naukkis]

Anyone that thinks Zen 3 will beat Alder Lake in anything needs to go to AT bench, pick a ST benchmark of the 11900k, and add 15-20% to it, and see which number (Zen 3 or the new number) is higher. Hint: Alder Lake will be in most scenarios.

For gaming more cache does much more than higher IPC. That AMD's +15% uplift is from flat 4GHz comparison, for actual products at near 5GHz that percentage is higher. And those titles which don't scale as they fit already Zen3's L3 Zen3 has something like 50% performance gap between Rocketlake and Zen3. Alderlake doubles L3 against Rocketlake but it's still far cry from being able to take gaming crown away from AMD's monster cache gaming CPU's.

 

[gk1951]

Dr. Lisa Su has instilled a sense of keeping your mouth shut until you have parts that work. AMD has been very quiet during the Alder Lake presentation.

I think they have an answer waiting in the wings but want to make sure it works.

 

[eek2121]

Really? I must have missed that part.

As secretive as AMD is about everything, to the point of describing last years parts at conferences, not upcoming ones, why do you think AMD spilled everything about their V-Cache at the demo, left nothing for the launch?

This is the slide I saw from their Hotchips presentation. Is there a different one I'm supposed to be looking for?

I thought I saw it on the liveblog. I will see if I can find it.

 

[Hitman928]

I thought I saw it on the liveblog. I will see if I can find it.

If I may, you're probably getting confused from this post where @Asterox posted a video from the original Zen 3 launch where they mentioned 72 MB of "game cache" (just o.g. Zen 3 L3 + L2 max combined total) and in the same post, included a picture of the Zen3d gaming benchmarks from AMD. I believe the intent was to show how the "3D V-cache" was an extension of AMD's original "game cache" marketing, but the "3D V-cache" prototype was clearly 192 MB L3 (32 MB bases + 64 MB stack per CCD) from their Hot Chips presentation.

 

[Timorous]

I agree. The numbers will likely be all over the place, but the gaming SKU Intel is launching Alder Lake with, 12900k will be convincingly higher than 11900K and Zen 3.

This Geegbench score analyzed by VideoCardz looks strong.
Intel Core i7-12700 non-K Alder Lake CPU is almost as fast as Ryzen 7 5800X in leaked Geekbench score - VideoCardz.com

The FP and INT scores are not that strong vs the 5800X. Looking at some averages INT may be around 10% faster for this 12700 vs 5800X but FP is looking about the same give or take.

If I had to guess I would say that in reviews where they test 1080p max the performance difference between the 12900K and the 5900X in gaming will be pretty small, probably sub 5%.

The low res and low settings tests that Ian does will be more interesting to spot differences and here it might draw level with the 5900X/5800X. Looking at those tests ADL needs to improve by 20% + in some games to match Zen 3 at the more CPU limited settings and at more GPU limited settings your CPU does not matter as much.

 

[eek2121]

The FP and INT scores are not that strong vs the 5800X. Looking at some averages INT may be around 10% faster for this 12700 vs 5800X but FP is looking about the same give or take.

If I had to guess I would say that in reviews where they test 1080p max the performance difference between the 12900K and the 5900X in gaming will be pretty small, probably sub 5%.

The low res and low settings tests that Ian does will be more interesting to spot differences and here it might draw level with the 5900X/5800X. Looking at those tests ADL needs to improve by 20% + in some games to match Zen 3 at the more CPU limited settings and at more GPU limited settings your CPU does not matter as much.

Intel Corporation Alder Lake Client Platform - Geekbench Browser

Benchmark results for an Intel Corporation Alder Lake Client Platform with an Intel Core i9-12900K processor.

browser.geekbench.com

1893/17299
That is the 12900k. Maybe AMD will drop some more definite info soon. Competition is grand.

 

[Asterox]

If I may, you're probably getting confused from this post where @Asterox posted a video from the original Zen 3 launch where they mentioned 72 MB of "game cache" (just o.g. Zen 3 L3 + L2 max combined total) and in the same post, included a picture of the Zen3d gaming benchmarks from AMD. I believe the intent was to show how the "3D V-cache" was an extension of AMD's original "game cache" marketing, but the "3D V-cache" prototype was clearly 192 MB L3 (32 MB bases + 64 MB stack per CCD) from their Hot Chips presentation.

That or this video, this is from 2019(July 1) or Zen 2/3000 lounch.

Zen 2/R5 3600/32mb L3 cache vs Zen 2/R5 4650G/8mb L3 cache. As we now, in 90% of situations big L3 cache can offer beeter CPU performanse vs R5 4650G, in gaming especially."But if you look in the details", R5 4650G/AMD VCN hardware speeds up CPU video encoding and decoding by 50-100%, depending on the application used.It would be very useful, if R5 3600 has integrated VCN hardware.Finally the question of price = performance, in my country R5 3600 =R5 4650G same price 200euro.

AMD 3D V-Cache, "it is very simple if AMD intends to put 64mb on each CCD."

R5 5600XT/32mb L3 cache

- one CCD, 64mb 3D V-Cache

R5 5800XT/32mb L3 cache

- one CCD, 64mb 3D V-Cache

R9 5900XT/64mb L3 Cache

- two CCD-s 2 x 64mb 3D V-Cache

 

[LightningZ71]

Remember, it's possible that AMD will still use die harvesting on lower end parts. They'll know that a given 5600x has disabled cores before the bonding process, and they'll know that some SRAM cells will be bad in some SRAM die before bonding. They could choose to bond 48MB or 32MB SRAM dies to 5600X dies to make 5600x parts with a total of 64 or 72 MB of L3.

However, I don't expect AMD to produce a 5600X3D

 

[Mopetar]

I dunno, I still believe it'll be 6000 series. Rembrandt will be included in that.

Compared to previous generations only half of the stack has been released for the 5000 series. It makes far more sense for a single one off product with a weird name like "5000 XT" than making an entire new series which will probably have even fewer products in it.

Releasing a 6000 series makes everyone think your 5000 series is obsolete. No need to Osborne your own product line just for the sake a small number of high-end CPUs most people can't afford that AMD will have even more problems supplying.

 

[jpiniero]

Compared to previous generations only half of the stack has been released for the 5000 series. It makes far more sense for a single one off product with a weird name like "5000 XT" than making an entire new series which will probably have even fewer products in it.

Releasing a 6000 series makes everyone think your 5000 series is obsolete. No need to Osborne your own product line just for the sake a small number of high-end CPUs most people can't afford that AMD will have even more problems supplying.

AMD would make new SKUs for the non vcache models with small frequency tweaks.

I was thinking something like this:

6950VX
6900VX
6900 (OEM only)
6800X
6800 (OEM only)
6600X

Then throw in Rembrandt in there.

 

[moinmoin]

Releasing a 6000 series makes everyone think your 5000 series is obsolete.

Just like releasing the 4000 series made the 3000 series obsolete. AMD will increase the number as it wants, there's no logic required to it.

 

[Mopetar]

Just like releasing the 4000 series made the 3000 series obsolete. AMD will increase the number as it wants, there's no logic required to it.

There's a reason that AMD moved to unify the numbering on their product lines.

Maybe it's not a problem for people on tech forums, but the average consumer is practically clueless and something as simple as a numbering scheme is a big deal to them.

 

[xilli_fiberbit]

MILAN-X seems to be already ready ?

https://twitter.com/x/status/1430551880824803329

So first MILAN-X and then VERMEER-X ?

 

[jpiniero]

MILAN-X seems to be already ready ?

https://twitter.com/x/status/1430551880824803329

So first MILAN-X and then VERMEER-X ?

Probably not ready, just that they are starting final production. If Zen 3D was going to launch at CES you'd need to start pretty soon.

 

[Joe NYC]

Remember, it's possible that AMD will still use die harvesting on lower end parts. They'll know that a given 5600x has disabled cores before the bonding process, and they'll know that some SRAM cells will be bad in some SRAM die before bonding. They could choose to bond 48MB or 32MB SRAM dies to 5600X dies to make 5600x parts with a total of 64 or 72 MB of L3.

However, I don't expect AMD to produce a 5600X3D

The SRAM probably has enough spare capacity to still allow full 64MB with a small defect. Creating a separate SKU based on lower than 64 MB V-Cache may just be too many SKUs.

Also, if 5600x may not be a very good die to even bother adding V-Cache to. There are tons of better dies to enhanced with V-Cache

 

[Joe NYC]

Probably not ready, just that they are starting final production. If Zen 3D was going to launch at CES you'd need to start pretty soon.

Unless this is the new "B2" stepping that needs the entire fabrication process from start, no, if it is just standard Zen 3 die, turning it into Zen 3D, the lead time is limited.

~2 weeks? Certainly not 2 full quarters.

 

[jpiniero]

Unless this is the new "B2" stepping that needs the entire fabrication process from start, no, if it is just standard Zen 3 die, turning it into Zen 3D, the lead time is limited.

~2 weeks? Certainly not 2 full quarters.

They might do a new stepping but you also have to fab the cache die. I have no idea how long the cache die would take. Then you have the bonding process.