Discussion Speculation: Zen 4 (EPYC 4 "Genoa", Ryzen 7000, etc.)

[Vattila]

Except for the details about the improvements in the microarchitecture, we now know pretty well what to expect with Zen 3.

The leaked presentation by AMD Senior Manager Martin Hilgeman shows that EPYC 3 "Milan" will, as promised and expected, reuse the current platform (SP3), and the system architecture and packaging looks to be the same, with the same 9-die chiplet design and the same maximum core and thread-count (no SMT-4, contrary to rumour). The biggest change revealed so far is the enlargement of the compute complex from 4 cores to 8 cores, all sharing a larger L3 cache ("32+ MB", likely to double to 64 MB, I think).

Hilgeman's slides did also show that EPYC 4 "Genoa" is in the definition phase (or was at the time of the presentation in September, at least), and will come with a new platform (SP5), with new memory support (likely DDR5).

What else do you think we will see with Zen 4? PCI-Express 5 support? Increased core-count? 4-way SMT? New packaging (interposer, 2.5D, 3D)? Integrated memory on package (HBM)?

Vote in the poll and share your thoughts!

 

[Exist50]

For those of you wanting to build a Micro-ATX Genoa Workstation

View attachment 71617

https://www.asrockrack.com/general/productdetail.asp?Model=GENOAD8UD-2T/X550

"Deep" mATX, so probably won't fit in many typical mATX cases. But I love that Asrock is carrying on the tradition of cramming absurd amounts of hardware into tiny footprints.

 

[nicalandia]

"Deep" mATX, so probably won't fit in many typical mATX cases. But I love that Asrock is carrying on the tradition of cramming absurd amounts of hardware into tiny footprints.

Think about this as an investment. Zen4/Zen4X3D/Zen4c Will likely fit on that tiny package.

 

[nicalandia]

That Micro-ATX has 8 DIMM.. If you want more Gigabyte has you covered with this 48 DIMM solution for two socket Genoa System

 

[biostud]

In a world where the RTX 4090 is being torn off the shelves, I'm sure a 7950X3D would sell as well.

But for most of us we would probably choose a more price conscious product.

 

[desrever]

AMD will not release a 16 V Cache because a second chiplet will not give more gaming performance than the single CCD and they will be selling each of those at top dollar as in Genoa-X

Each chiplet costs peanuts for AMD. Selling any higher end chip is better and higher margin. It would make more sense to only have 16 core than not have it. Why would they sell a 7800X3D for half as much as a 7950X3D. The cost to AMD to make the 7950X3D is probably like $80 and the cost of the 7800X3D for AMD is probably like $60. They would make much more money selling an $800 vs a $400 one.

If the argument is for using them all in EPYC then there would be no reason to launch a 7800X3D. The only reason they didn't launch a 5950X is probably cause then there would barely be any reason for anyone to buy AM5 and Zen4.

 

[nicalandia]

AMD 7950X 16C vs Apple M1 Ultra 20 Cores

MacPro7,1 vs Mac Studio - Geekbench

browser.geekbench.com

 

[gdansk]

It seems Hackintosh isn't really great for performance. For reference that's an even lower score than I get in Windows with my air-cooled, stock 7950X (2225/24097).
Also I can't help but laugh at the AES scores. Two massacres but in opposite directions. Geekbench 6 should remove that.

 

[jamescox]

Gamers Nexus nexus had an interview with an AMD engineer about RDNA3. They discussed some of the cost savings of the chiplet architecture. I had not seen the info about ‘infinity fan out links’; I guess we might not get much stacking technology since they seem to be coming up with cheaper alternatives. Although, this requires that the chips be very close, so SerDes GMI may be around a bit longer.

 

[APU_Fusion]

Is there any hint for when the 7950x3d will be available? Going to upgrade from 4790k and 1070ti to 4090. Getting itchy.

 

[In2Photos]

Is there any hint for when the 7950x3d will be available? Going to upgrade from 4790k and 1070ti to 4090. Getting itchy.

X3D versions are rumored to be announced at CES in January. But there are no guarantees which models, with several speculating that there will only be a 6-core and 8-core variant.

 

[Mopetar]

I think there is still a chance we'll see 8C and 16C Vcache models. I'm basing this on the original AMD presentation that had 5900X3D (which required two Vcache addon chiplets). We'll know soon.

I remain doubtful about a 16C model. There may be a few instances where the extra cache boosts performance for an application, but as we saw with the 5800X3D, it was only really the gaming market that had that big performance gain.

When the margins for server parts are already better, AMD maximizes consumer exposure by selling more 8C parts. We probably wouldn't even see consumer v-cache if the average game didn't do so well with the extra cache. But being able to pretty easily have the top gaming CPU on the market is a niche well worth occupying even if your margins aren't as good.

 

[Joe NYC]

Is there any hint for when the 7950x3d will be available? Going to upgrade from 4790k and 1070ti to 4090. Getting itchy.

Not a word so far. AMD is good at keeping things close to the vest.

The general expectation is that it will be announced at CES (shortly after New Year) and then availability sometimes afterwards. Which is the big question - how long afterwards.

I am waiting as well, for the 8 core variant...

 

[Joe NYC]

I remain doubtful about a 16C model. There may be a few instances where the extra cache boosts performance for an application, but as we saw with the 5800X3D, it was only really the gaming market that had that big performance gain.

When the margins for server parts are already better, AMD maximizes consumer exposure by selling more 8C parts. We probably wouldn't even see consumer v-cache if the average game didn't do so well with the extra cache. But being able to pretty easily have the top gaming CPU on the market is a niche well worth occupying even if your margins aren't as good.

A lot of people are going with the assumption that the V-Cache will be N6. If it is the case, the capacity for the cache alone is endless. AMD could order on for every single Zen 4 CCD and barely scratch the surface of TSMC available capacity.

There may be a capacity constraint on assembly, but this is area where TSMC has been investing into heavily.

My wild guess is that there will be a V-Cache version of every chip, 6, 8, 12, 16 core, and AMD will likely price them at the original into-MSRP of the same chips without V-Cache.

 

[Mopetar]

A lot of people are going with the assumption that the V-Cache will be N6. If it is the case, the capacity for the cache alone is endless. AMD could order on for every single Zen 4 CCD and barely scratch the surface of TSMC available capacity.

There may be a capacity constraint on assembly, but this is area where TSMC has been investing into heavily.

My wild guess is that there will be a V-Cache version of every chip, 6, 8, 12, 16 core, and AMD will likely price them at the original into-MSRP of the same chips without V-Cache.

But what's going to make consumers purchase them? The only thing they do consistently better is gaming and you don't need more than 8 cores for most games on the market and in some cases the split chiplet design hurts performance.

AMD is already having a hard time moving new Zen 4 chips. Adding another $100 on top of that for now benefit outside of gaming won't help sell to anyone but gamers.

I think the real possible benefit is in their APUs. Extra v-cache or using some as an infinity cache for the GPU cores would provide some substantial performance uplifts and help compensate for the lower memory bandwidth since both CPU and GPU cores are using the same system memory and bus. It also let's AMD make a smaller base die if cache can be shifted to a vertical layer.

 

[Markfw]

But what's going to make consumers purchase them? The only thing they do consistently better is gaming and you don't need more than 8 cores for most games on the market and in some cases the split chiplet design hurts performance.

AMD is already having a hard time moving new Zen 4 chips. Adding another $100 on top of that for now benefit outside of gaming won't help sell to anyone but gamers.

I think the real possible benefit is in their APUs. Extra v-cache or using some as an infinity cache for the GPU cores would provide some substantial performance uplifts and help compensate for the lower memory bandwidth since both CPU and GPU cores are using the same system memory and bus. It also let's AMD make a smaller base die if cache can be shifted to a vertical layer.

I would buy them. They probably would help the same applications that make Milan-x so great.

 

[Mopetar]

I would buy them. They probably would help the same applications that make Milan-x so great.

Your use case is certainly different than most around here, but did you ever run any of you workloads on Zen3D to see if they benefit at all.

If the voltage issue still exists, most applications would regress when Zen 4 runs full out as the v-cache doesn't make up for the clock speed loss outside of gaming for the most part.

I recall a few applications that saw performance gains from v-cache when I had looked into it myself some months ago, but they were outliers. Why spend an extra $100 if it does nothing for you?

 

[Markfw]

Your use case is certainly different than most around here, but did you ever run any of you workloads on Zen3D to see if they benefit at all.

If the voltage issue still exists, most applications would regress when Zen 4 runs full out as the v-cache doesn't make up for the clock speed loss outside of gaming for the most part.

I recall a few applications that saw performance gains from v-cache when I had looked into it myself some months ago, but they were outliers. Why spend an extra $100 if it does nothing for you?

there are several apps that gain. The last primegrid we had, its you use a process to keep the thread in one ccd, the l3 fit the entire task, and run times were greatly reduced. So, yes, I know that I would buy a 7950x3d. I know not many would. I want a Milan-x, but the cheapest like like $4500 or so. Out of my budget. Edit, and I would pay $200-300 more, like $800-900 for a 7950x3d. All of the DC forum would. And all the other DC groups would (thats a lot of groups)

 

[Joe NYC]

But what's going to make consumers purchase them? The only thing they do consistently better is gaming and you don't need more than 8 cores for most games on the market and in some cases the split chiplet design hurts performance.

By being one (or two) steps ahead of Intel at every task you want to target.

8 core is all that 95+% consumers need, but for those who could use extra cores, they can get it (with 7950x3d) without losing gaming performance of 8 core V-Cache processors.

AMD is already having a hard time moving new Zen 4 chips. Adding another $100 on top of that for now benefit outside of gaming won't help sell to anyone but gamers.

I think AMD should make the current price cuts for non-V-Cache versions permanent, and introduce the V-Cache versions at the original MSRP.

For example, 8 core Zen 4 with V-Cache for $399 would be a steal. I am not suggesting discontinuing the non-V-Cache models, there would be some small savings for those...

I think the real possible benefit is in their APUs. Extra v-cache or using some as an infinity cache for the GPU cores would provide some substantial performance uplifts and help compensate for the lower memory bandwidth since both CPU and GPU cores are using the same system memory and bus. It also let's AMD make a smaller base die if cache can be shifted to a vertical layer.

There are likely some technical issues to overcome for some sort of L3 sharing between CPU and GPU. I am not sure if AMD was able to solve it. It might imply unified memory - something that would need Microsoft to make it work in Windows.

Somehow, I don't think we will get there with this upcoming version of APUs. The only possibility I see, if there is some kind of stacked cache, it will not be shared between CPU and GPU.

It may still be a good idea, as you say, to keep the base die (N5) smaller, and possibly add L3 on cheaper N6 node.

But in 2023, AMD should be introducing Mi300 that will have unified memory...

 

[LightningZ71]

It seems having the iGPU share the L3 isn't something that AMD has figured out yet in their APUs. It would seem to me that they could get good mileage out of an APU configured to have an 8MB L3 on the base die in the CCX and an additional 32MB on a stacked L3 die. Then, implement a 16MB Infinity cache for the iGPU on the base die with it. That only adds an additional 8MB of base die SRAM over the 6000 series APUs and allows peoduct segmentation between having the stacked L3 for the CCX and for enabled infinity cache (it could be arranged so that a half can be disabled for yield).

The stack would look like:
8/16, 40MB l3, full iGPU with 16MB IFc
6/12, 40MB (maybe 24MB if 1/2 cache die disabled) l3, 2/3 iGPU with 8MB IFc
4/8, 8MB L3, 1/3 iGPU w/ 0MB IFc

 

[Mopetar]

By being one (or two) steps ahead of Intel at every task you want to target.

8 core is all that 95+% consumers need, but for those who could use extra cores, they can get it (with 7950x3d) without losing gaming performance of 8 core V-Cache processors.

Why buy a 7950X3D instead of a 7950X though? AMD has to compete against their own products as well and you can look at productivity benchmarks for Zen3D and see for yourself that in most applications outside of gaming, the extra cache doesn't do anything.

I'd like the 5800X3D it winds up having lower clock speeds, in some cases it will perform worse than a 7950X. The market for people who'll buy an expensive CPU for productivity but sacrifice that so they can game a bit better is pretty slim.

 

[DrMrLordX]

Is there any hint for when the 7950x3d will be available? Going to upgrade from 4790k and 1070ti to 4090. Getting itchy.

Based on estimates of when 3d cache stacking is/would be ready this year and some other commentary, earliest time they would release it would be feb/march of 2023.

Why buy a 7950X3D instead of a 7950X though?

Halo part, game streaming, stuff like that. Why buy a 12900KS? People still buy those things. Intel will launch a 13900KS more likely than not. It should be easy enough for AMD to add the SKU and produce a part that can win all the benchmarks.

 

[Joe NYC]

Why buy a 7950X3D instead of a 7950X though? AMD has to compete against their own products as well and you can look at productivity benchmarks for Zen3D and see for yourself that in most applications outside of gaming, the extra cache doesn't do anything.

I think it is going to be a little different with Zen 4 V-Cache, because most likely, you will not be starting with ~5 to 12% clock speed deficit. So, I think nearly every single application will get some benefit, even if it is low single digit percentage.

Why get 7950x3d instead of 7950x? Getting either one means you want or need 16 cores. Getting 7950x3d means you are not losing out on gaming performance for going to 16 cores. Meaning, you can have both (if you are willing to pay), you don't have to chose one or the other.

AMD cost for adding 1 layer of V-Cache may be $10 (just a guess) so $20 for two. So I don't think pricing difference will be terribly steep. Perhaps +$50 on the retail price per the 64 MB V-Cache.

Right now, the difference between 5800x and 5800x3d is $80. I think AMD may be able to reduce the difference as the V-Cache really hits mass production.

And, perhaps this time, Microsoft Azure will not vacuum the entire production, as it happened with Milan-X, leaving nothing for the gamers, for months. This time, AMD is apparently planning on launching Genoa-X in H2 2023, so hopefully, there will be plenty of supply in H1 2023 for the consumer side of Zen 4 (which can use some help after the shaky launch).

 

[biostud]

7950X @ $549 Amazon

Amazon.com: AMD Ryzen™ 9 7950X 16-Core, 32-Thread Unlocked Desktop Processor : Everything Else

 

[igor_kavinski]

I hope this signals a revised, better performing silicon coming soon once they clear out existing Zen 4 stock.

 

[eek2121]

Why buy a 7950X3D instead of a 7950X though? AMD has to compete against their own products as well and you can look at productivity benchmarks for Zen3D and see for yourself that in most applications outside of gaming, the extra cache doesn't do anything.

I'd like the 5800X3D it winds up having lower clock speeds, in some cases it will perform worse than a 7950X. The market for people who'll buy an expensive CPU for productivity but sacrifice that so they can game a bit better is pretty slim.

There is a very good possibility that clocks won’t be lower this gen. Clocks were lower on the previous generation due to needing to keep the voltage in check. The reduction in clocks masked some of the performance gains in some applications.