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"

 

[andermans]

Just do LLVM or Chromium (or Android .... Their build system is pretty parallel), should be easy enough and pretty much just compilation. Very curious how well the extra cache works for that

 

[Joe NYC]

I still maintain that AMD should aim for a dual product launch, with Zen3D/Vermeer-X supporting those still on AM4 and Zen4/Raphael supporting those looking for a new platform with (overall) cutting-edge performance. There may be some odd instances where Zen3D will outperform Zen4, but not very many.

Yup, especially since these 2 are coming from 2 different buckets (N7 vs. N5)

 

[Joe NYC]

I would consider a Zen 3D new build depending on benchmarks though. The massive caches will perform exceptionally well for some applications. They should do really well for code compile machines. I suspect the we are going to need a bigger benchmark for Milan-x. The Linux kernel compile was already down to 20 seconds or so for some high end Epyc systems.

It's possible that the new stepping will enable a slight frequency increase, in addition to cache. We will see...

Either Zen3 / Zen 3D or Alder Lake, it is not an investment in a long term platform. Alder lake has 1 useless upgrade before switching to Meteor Lake, which will be a new platform.

So from upgradability, it is a wash. Socket AM5 will have a long lifespan potential, but will be a while before it comes out, and there may still be DDR5 premium even a year from now...

 

[Insert_Nickname]

Socket AM5 will have a long lifespan potential, but will be a while before it comes out, and there may still be DDR5 premium even a year from now...

Indeed. But we might see the same segmentation as on AM4 in that early boards may not be able to run the later generation APU/CPUs. So it might be better to skip the first generation boards, kind of how the B450 ended up as the most compatible AM4 chipset. It can take everything from BR to Vermeer on most boards.

 

[coercitiv]

I don't think this is true. AMD never had an overall performance lead over Intel until Zen 3 and they sold fine even when they were further behind.

If Zen3D didn't happen at all I don't think AMD would be any worse off if they had nothing new until Zen 4.

They could not charge a premium however.

AMD with Zen3 vs. Golden Cove is in a much better shape than in any other previous period:

• this time around they start with a solid brand image, it's Intel that needs to punch through the marketing wall
• their CPU premium is compensated by the very high Z690 pricing, absurdly high if you count DDR5
• Alder Lake suffers from a combination of factors that make the platform less appealing to some customers: Win11 perception, early E-core compatibility problems, socket compatibility for coolers, motherboard availability. Many of these may seem a non-problems for some of you, but in some regions such as mine they were real. (think next to zero availability for LGA 1700 compatible coolers at launch - good luck powering your system as an average consumer)

All of these factors decay over time, but for now AMD can still sell at a premium while they prepare the 3D stock and make sure they match Intel in the eyes of the consumer early 2022. In my EU country I have seen no sign of AMD lowering prices, the only change I saw was a more steady supply. (as opposed to selling out)

Are people really going to make a new build with Zen 3D on a platform that's got to future beyond that as opposed to going with Intel when you can get in at the start of one? AMD fans would hold out for Zen 4 anyways especially with the GPU market still being the way it is.

The same can be said for every 1 out of 2 Intel generations. As long as the feature set of late AM4 boards is solid enough, users looking for flagship performance or users upgrading from older systems won't care that much.

 

[igor_kavinski]

Aren't there a lot of other bottlenecks that way?

The storage I/O bottleneck can be removed by using a ramdrive. Then it's just the CPU/memory subsystem being stressed.

 

[DrMrLordX]

DIY market is not big enough for something like that.

AMD can always reduce volume. We don't know how many of the Genoa CCDs will test well enough to go in actual Genoa products. Early Raphael can come from those CCDs. Plus, unlike Vermeer-X/Zen3D, Raphael need not be DiY-only.

That AMD is doing Zen 3D shows that Raphael is a long ways off. Might take AMD months to fill the Genoa Cloud orders.

That's a possibility, and I admit that I am also very pessimistic about AMD prioritizing consumer Zen4 products, but at the same time . . . not all of those CCDs can go in Genoa. The more Genoa they try to produce, the more leftovers they get that have too-high leakage for server duty.

 

[moinmoin]

AMD can always reduce volume.

Not really, at least not the way you appear to suggest there. The lead times are significant, and we complained about lack of supply before which AMD couldn't resolve for a long time due to the lead times required for any course changes. And the pandemic likely exacerbated the need to plan significantly ahead to ensure a steady stream of supply.

 

[DrMrLordX]

Not really, at least not the way you appear to suggest there.

How so? They can sit on dice, or just reduce DiY volume by pushing Raphael into OEM channels. There's also Raphael-H which may (or may not) use some of the same core components as desktop Raphael. There should be plenty of room for N7, N6, and N5 products to coexist in AMD's lineup. We already know that AMD will continue selling Vermeer as B2-stepping 5-series chips well into 2022, overlapping at least with Zen3d. I suspect Raphael volumes will be low-to-non-existent for several months post-launch anyway. The only way that doesn't happen is if they stockpile dice well in advance of a delayed launch.

 

[Joe NYC]

AMD can always reduce volume. We don't know how many of the Genoa CCDs will test well enough to go in actual Genoa products. Early Raphael can come from those CCDs. Plus, unlike Vermeer-X/Zen3D, Raphael need not be DiY-only.

That's a possibility, and I admit that I am also very pessimistic about AMD prioritizing consumer Zen4 products, but at the same time . . . not all of those CCDs can go in Genoa. The more Genoa they try to produce, the more leftovers they get that have too-high leakage for server duty.

Depends on how the SKUs end up being configured. If AMD is going to stick with offering full (8-12) CCD, low core count, like the "F" series, they should be able to harvest one very good core from just about every CCD.

 

[biostud]

Indeed. But we might see the same segmentation as on AM4 in that early boards may not be able to run the later generation APU/CPUs. So it might be better to skip the first generation boards, kind of how the B450 ended up as the most compatible AM4 chipset. It can take everything from BR to Vermeer on most boards.

But wasn't one of the main reasons for the lack of support a too small ROM chip, that couldn't hold the data needed for all the different generations of CPUs? And the later generation of motherboards had doubled the capacity of the ROM.

 

[moinmoin]

How so? They can sit on dice, or just reduce DiY volume by pushing Raphael into OEM channels. There's also Raphael-H which may (or may not) use some of the same core components as desktop Raphael. There should be plenty of room for N7, N6, and N5 products to coexist in AMD's lineup. We already know that AMD will continue selling Vermeer as B2-stepping 5-series chips well into 2022, overlapping at least with Zen3d. I suspect Raphael volumes will be low-to-non-existent for several months post-launch anyway. The only way that doesn't happen is if they stockpile dice well in advance of a delayed launch.

You can't use dice as is, they still need to be assembled, packaged and shipped. So you not only have the lead time at foundries but also for everything that follows until the resulting products are actually sold and delivered to some customers.

 

[Topweasel]

But wasn't one of the main reasons for the lack of support a too small ROM chip, that couldn't hold the data needed for all the different generations of CPUs? And the later generation of motherboards had doubled the capacity of the ROM.

AMD learned some hard lessons with AM4. We think of Retail purchases as a large source of income. But realistically OEM sales drives sales and profitability. AM4 was a compromised socket driven completely by OEM demands. They wanted a single platform for APU's and Desktop chips, they wanted Bulldozer support while they waited for Raven Ridge launch, and they didn't want it to be LGA. That had its own downsides on its own. But also they ran into the issue with the retail board companies cutting corners. Their goals in letting the platform having legs that last was more on the manufacturers. They wouldn't have to start over with a new board for every gen. Intel would probably change platforms every generation if it wasn't for them agreeing to a 2 year cycle with the OEM's. AMD wanted to promise them longer.

So the question going forward is will AMD let themselves get in that trap again. It could go either way. They could have tighter controls over the implementation on retail boards or they could decide its easier to follow Intel's process and not get pushed into a corner where retail partners are making decisions that impact their image. The move to PCIe 4 also is another point. PCIe 5, DDR5, and USB/TB4 should have long legs, but do they want to get pushed into another scenario where they might have to move slowly because they don't want to change the platform. AMD was trying to hit a middle ground between the needless amount of platform changes Intel made and not having any technology changes for years because of the requests of the retail market. It goes beyond not having enough memory on the BIOS. In the end I doubt they model their behavior off of what they did with AM4.

 

[Topweasel]

You can't use dice as is, they still need to be assembled, packaged and shipped. So you not only have the lead time at foundries but also for everything that follows until the resulting products are actually sold and delivered to some customers.

Yeah I think the big issue is we don't fully understand the process. How the wafer orders are made, how they get split up at the lab, how they are getting packaged onto the substrate. It could be difficult as they ask for X wafers for Genoa production, all working chips get installed on Genoa substrates. Or it could be a bin of dies and they say ooh this one is good for Genoa and ship it off for installation. Or its something in between. I doubt its as easy as calling up one dept or company and going hey we need you to shift X more dies to Y product line.

 

[zir_blazer]

But wasn't one of the main reasons for the lack of support a too small ROM chip, that couldn't hold the data needed for all the different generations of CPUs? And the later generation of motherboards had doubled the capacity of the ROM.

It goes beyond not having enough memory on the BIOS.

The entire ROM size thing was proven to be pretty much a lame, bogus technical reason to justify not supporting newer Processors. It was never a real problem to begin with, just an excuse.

 

[Topweasel]

The entire ROM size thing was proven to be pretty much a lame, bogus technical reason to justify not supporting newer Processors. It was never a real problem to begin with, just an excuse.

Pretty much. AMD released a new platform with new functions and performance envelopes, like all companies AMD wants to make sure their products are used in their performative setting and without surprises that even if not their fault they don't have their rep tarnished by another party (whether they are a partner or not). Heck just in terms what Intel can get away with giving them the ammo for a lopsided slide can be bad enough.

Intel for multiple reason's has to be clear in the PR materials where and how they sourced their performance metrics (now they don't even have to have it in the slide but an embedded link). This means that the public is free to critique and generally Intel gets slapped if they twist the truth to much (like recompiling a benchmark with their compiler that based on whether it has genuine Intel tag or not it sends it down a performance path or one without many simd's. Letting the public run a 5950X on a B350 board, with thermal throttling limits, no PCIe 4, no direct memory access, and so on, the numbers would look so much worse and no amount of "but look at the footer webpaged. They rigged it to fail" is going to help as barely any body gets that far or cares enough at the end of the point.

Now look at Alderlake. Thread director meant almost all of the numbers for legitimacy needed to be taken from Windows 11. That meant any AMD comparisons had to be done in Windows 11. Then after the fact we are told about a bug in the scheduler that only affects AMD processors exists. Doesn't matter Intel now has the slides with crappy 11 game performance that will use till Raptor Lake.

I like using these because its just a perfect example of their direct competitor using a working configuration against them. But that's the point. It's why AMD and Intel both pushed away chipset partners. It's why NVidia puts so much into their FE cards when the cooling alone would cost the AIB's amost as much as the MSRP that Nvidia sets and will only let the initial reviews be on the FE models. Its about controlling the representation of the product. 300 series boards especially were all sorts of half baked and AMD didn't want their newer super competitive CPU's to be represented on those boards.

 

[Insert_Nickname]

But wasn't one of the main reasons for the lack of support a too small ROM chip, that couldn't hold the data needed for all the different generations of CPUs? And the later generation of motherboards had doubled the capacity of the ROM.

That was the official version at least. The latest UEFI version on my A320M-K does remove BR support, but adds Ryzen 5000-series support. So I'm betting that is a trade-off most DIYers can live with.

 

[Shivansps]

But wasn't one of the main reasons for the lack of support a too small ROM chip, that couldn't hold the data needed for all the different generations of CPUs? And the later generation of motherboards had doubled the capacity of the ROM.

That was a lie to sell new motherboards. Plain and simple.

A B450M-DS3H with a 16MB ROM can support everything from the Ryzen 1000 to the 5950X, incluiding all Athlons and APUs minus BR. On A320 is now the same as well with the difference that the Agesa is blocking Cezanne from booting up, but the microcode is in there.

Rom size was never a problem.

 

[jpiniero]

AMD learned some hard lessons with AM4. We think of Retail purchases as a large source of income. But realistically OEM sales drives sales and profitability. AM4 was a compromised socket driven completely by OEM demands. They wanted a single platform for APU's and Desktop chips, they wanted Bulldozer support while they waited for Raven Ridge launch, and they didn't want it to be LGA. That had its own downsides on its own. But also they ran into the issue with the retail board companies cutting corners. Their goals in letting the platform having legs that last was more on the manufacturers. They wouldn't have to start over with a new board for every gen. Intel would probably change platforms every generation if it wasn't for them agreeing to a 2 year cycle with the OEM's. AMD wanted to promise them longer.

But if you think about it, Intel didn't really change the physical socket much for awhile. Intel went from 1156 to 1155 to 1150 to 1151 pins. AMD could have done the same thing with AM4, change a pin or two and call it a new socket.

 

[biostud]

That was a lie to sell new motherboards. Plain and simple.

A B450M-DS3H with a 16MB ROM can support everything from the Ryzen 1000 to the 5950X, incluiding all Athlons and APUs minus BR. On A320 is now the same as well with the difference that the Agesa is blocking Cezanne from booting up, but the microcode is in there.

Rom size was never a problem.

So based on this, then there shouldn't be a problem getting first generation AM5 boards.

 

[Topweasel]

But if you think about it, Intel didn't really change the physical socket much for awhile. Intel went from 1156 to 1155 to 1150 to 1151 pins. AMD could have done the same thing with AM4, change a pin or two and call it a new socket.

That is exactly the point at some points they would need to change the platform. The move to DDR4. Almost everything else was not driven by CPU features and could probably could have used the same socket completely (with them claiming power requirements for the post Kaby-lake changes). They could do that. They may do that now. But they were trying to give their OEM's solutions that would allow them use their board designs for several generations. They were trying to through every bonus they could their way because they new they needed their support to survive. You only need to look at their board partners response when Zeppelin launched to see how their partners were treating them. I think one blamed AMD for their problems because they brought forward the release a couple of weeks because 1. It was around the Chinese New Year(legit) and 2. They had just spent all their effort launching I believe the 200 series Intel boards.

They were a second rate partner and supplier and no one had confidence in the product line and they wanted to make sure they did whatever they could to get OEM by in. What this meant for the retail market was second. I am sure they thought these companies wouldn't attach their names to complete crap and if they probably hadn't promised Dell, HP, and Lenovo that they would keep the platform going, 500 series would have been a new pin out. Now they have some clout and demand with these OEM's, I am guessing they will probably pick up a cycle closer to Intel. Expect rotating platforms. Specially if rumors of major Archs are going to get refreshes. Even if they don't, don't expect them to say anything about socket longevity, accept in past tense (how long AM4 or in the future AM5 lasted).

 

[Shivansps]

So based on this, then there shouldn't be a problem getting first generation AM5 boards.

No really, b350 AND x370 boards arent getting a bios update to support veermer because they are EOL.
A320 boards do get It only because they arent EOL AND there Is some unknown hidden reason for It.

I have heard that there are large amounts of unsold Renoirs PRO that they want to get rid off.

 

[Topweasel]

So based on this, then there shouldn't be a problem getting first generation AM5 boards.

Only if you assume that this past reason is the only reason AMD would have either a software or hardware compatibility with a board. I wouldn't assume its the case.

 

[eek2121]

AMD learned some hard lessons with AM4. We think of Retail purchases as a large source of income. But realistically OEM sales drives sales and profitability. AM4 was a compromised socket driven completely by OEM demands. They wanted a single platform for APU's and Desktop chips, they wanted Bulldozer support while they waited for Raven Ridge launch, and they didn't want it to be LGA. That had its own downsides on its own. But also they ran into the issue with the retail board companies cutting corners. Their goals in letting the platform having legs that last was more on the manufacturers. They wouldn't have to start over with a new board for every gen. Intel would probably change platforms every generation if it wasn't for them agreeing to a 2 year cycle with the OEM's. AMD wanted to promise them longer.

So the question going forward is will AMD let themselves get in that trap again. It could go either way. They could have tighter controls over the implementation on retail boards or they could decide its easier to follow Intel's process and not get pushed into a corner where retail partners are making decisions that impact their image. The move to PCIe 4 also is another point. PCIe 5, DDR5, and USB/TB4 should have long legs, but do they want to get pushed into another scenario where they might have to move slowly because they don't want to change the platform. AMD was trying to hit a middle ground between the needless amount of platform changes Intel made and not having any technology changes for years because of the requests of the retail market. It goes beyond not having enough memory on the BIOS. In the end I doubt they model their behavior off of what they did with AM4.

AMD just needs to make a top quality reference board. They can charge a premium for it so OEMs don’t get offended, however, they need to raise the bar on quality. No more of this day 1 broken bios stuff.

No really, b350 AND x370 boards arent getting a bios update to support veermer because they are EOL.
A320 boards do get It only because they arent EOL AND there Is some unknown hidden reason for It.

Well, the reason is actually pretty obvious.

 

[Mopetar]

The same can be said for every 1 out of 2 Intel generations. As long as the feature set of late AM4 boards is solid enough, users looking for flagship performance or users upgrading from older systems won't care that much.

I'd even go further and say this is the most exciting chip Intel has had in a long while, maybe even since Zen was released.

Trying to compare sales figures for this and draw conclusions about AMD needing Zen 3D are really overlooking the bigger picture and ignoring a lot of context.