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!

 

[trexfromouterspace]

That is Wrong. Zen4C is HIGH Density not lower..

It's Nosta, he's always wrong.

What are you on about? Is that sarcasm? How do you even cool something at 950W?

He's on meth.

 

[igor_kavinski]

He's on meth.

Maybe the last step in his meth cooking is adding some bits of silicon.

No offense to you, Nosta, by the way

 

[NostaSeronx]

What are you on about? Is that sarcasm? How do you even cool something at 950W?

Tachyum 128-core = >5.7 GHz at 950W TDP
It is targeting the same field of Cloud Native, but without the need of discrete Cloud-native GPUs(SIMD/SIMT accelerators) or TPUs(AI accelerators). TDP is basically CPU+GPU+AI in one package.
128-cores + 1 MB L2 per, no discrete L3, instead uses inactive L2 as L3.

128-core >5.7 GHz is liquid cooled <=== 950W
128-core >4 GHz is air cooled. <== 700W/600W

ARM hasn't released their cloud-native cores yet. However, they did drop some hints in a forum that they will be targeting higher clocks than 3 GHz with N2.

None of the high-end cloud-native processors are targeting low-clocks, none of them.

N5HPC => Higher frequency from reduced capacitance and resistance on both FEOL/BEOL with wider structures.
Zen4c => Smaller area from cut-down core.

Less area means less wire delay, and less stage delay, etc. Cut-down reduces total dynamic leakage meaning something that takes 230W PPT to achieve might only need less than 115W PPT to achieve.

 

[maddie]

It won't be a refresh, it will be a different feature set (more cache, lower frequencies) at a premium.

With the news that some of the limitations of present 3D cache being resolved (GHz ?), we might not even see a speed reduction to any significant degree.

 

[SteinFG]

Tachyum 128-core = >5.7 GHz at 950W TDP

Tachym is a scam lol, it's a Theranos of chip designers.

 

[Schmide]

Considering everything seems to be moving to OAM modules which at the moment top out around 600w. I highly doubt anything in the near future will come even close to 900w.

 

[Ajay]

It won't be a refresh, it will be a different feature set (more cache, lower frequencies) at a premium.

It's feature set is the same (functionally - same core), it'll deliver better gaming performance for more $$s because of the added cache. Maybe the old definitions are more fungible than before, given advanced manufacturing capabilities.

 

[NostaSeronx]

Tachym is a scam lol, it's a Theranos of chip designers.

It is just one of the known alternatives to Intel's monopoly of high frequency instances.

AmpereOneHF
AWS GravitonHF (High Frequency Neoverse cloud-native)
Alibaba Cloud T-headHF
etc.

The direction of Cloud-native development is towards High Frequency. Bergamo is launching in a swarm of 4~6 GHz cloud-native solutions.
Zen4 = 200W-400W
Zen4c = 320W-400W <-- 120W+ for 32 more smaller cores, yeah... Bergamo isn't going to be slow.

~Genoa lowest available TDP for 96-core ~200Ws (12x CCD @ ~16.7W)
~Bergamo lowest available TDP for 128-core ~320Ws (8x CCDs @ ~40W)

In everything I find, AMD is targeting the high-end of Cloud-native, and not the entry-end. This means AMD needs less supply to get a ROI and thus net profit.

 

[jamescox]

Zen4c is HPC, of which HPC has:
Wider poly pitch
Wider cell height
Wider metal pitch
Wider metal vias

The high density isn't from the libs, but rather from the macros/tiles.
Library = Less Dense
Architecture = More Dense

Same ISA;
Zen4 = Full 512-bit rate
Zen4c = Half 512-bit rate

Zen 4c does not seem like HPC. It is maximum density and compute per watt, which might be considered HPC in some circles. Genoa, likely with much higher max power consumption, will be the HPC part, or at least what I would call the HPC part. They also have the big APU for HPC, but probably not until later. Possibly before Grace-Hopper though.

I would expect Zen 4c to actually have a very good base clock due to the significantly lower power consumption. I suspect that it will have little to no boost clock though, so not good for mobile. You want a really high single core boost for mobile. In a throughput optimized situation, boost clock is often not useful. Many places actually just lock them in at base clock for performance consistency. They would generally be run at constant high load anyway. I am hoping that Zen4c is put together with silicon bridges and may utilize stacked L4 or L4 on the IO die.

 

[jamescox]

It won't be a refresh, it will be a different feature set (more cache, lower frequencies) at a premium.

They may have it available relatively quickly after Zen 4 release. There are a lot of things that seem to indicate that AMD cloud partners get really early access. Those cloud partners probably really want Zen 4 with v-cache. Who knows when we will get it in the consumer market though. The demand for Genoa is likely pretty high, but the server market moves slow and is often slow at switching to completely new platforms. Genoa will be all new boards, memory, and possibly chassis due to the increased power consumption. I have still seen Milan systems limited to lower parts rather than the the highest end Milan at 280 W. Anyway, with the larger L2, the large L3 may actually have a smaller effect in games than it did with Zen 3. The large L3 will likely still be very good for certain server / HPC applications.

 

[LightningZ71]

Zen4c will clock at least 2.5 Ghz no question about that. Also Perhaps Intel will enable HT on those Goldmont cores perhaps even SMT4, it only cost them about 10% die area for a 25%+ performance boost.

There's no point in adding HT to the 'mont cores. They are meant to be as space efficient as possible and are meant to optimize for single thread throughput in their space constraints.

I hope you are right for the 7950x vs 13900k matchup. By my own napkin math, if power and thermal constraints go out the window, the 13900k will lead most mt benchmarks by enough to make the point. Now, if you put a hard ceiling on power or temperatures, the story should be quite different.

 

[moinmoin]

It's feature set is the same (functionally - same core), it'll deliver better gaming performance for more $$s because of the added cache. Maybe the old definitions are more fungible than before, given advanced manufacturing capabilities.

For X3D models L3 cache is 3 times the size and people are going to have to pay a premium for that difference.

There are a lot of things that seem to indicate that AMD cloud partners get really early access.

Cloud partners, first and foremost Azure, always get first dips indeed. At FAD AMD indicated the cloud market is its single most important market by TAM:

(Source: slide 6 Dan McNamara presentation)

 

[biostud]

For X3D models L3 cache is 3 times the size and people are going to have to pay a premium for that difference.

Since they launched both the 5800X and 5800X3D @ $549, the difference doesn't have to be massive. Could be 7800X @ $499 and 7800X3D @ $599.

 

[randomhero]

Finally got time to watch FAD.

So, here goes my speculation.

I believe why Zen4 looks meh PPC wise is because AMD spent resources on new platforms and more important, they spent time on Zen4 being stepping stone to fully stacked devices in later generations of SoC.
I believe from next generation, Zen5, RDNA4, CDNA3, etc., all AMD SoCs to be stacked, except low end. So from mid to high end, every AMD product will include stacked silicon.

 

[deasd]

Sounds like Zen4 V-cache could be launched by the end of 2022?

https://twitter.com/x/status/1539171315940204544

https://twitter.com/x/status/1539179107090853888

 

[lobz]

Seems there is nowhere to run for intel. Don't misunderstand me, they're not dead, they absolutely do operate a very lucrative business, but... at the very end of 2022, AMD might be able to have a solution for every single use case that's better than what intel can offer.

 

[randomhero]

Sounds like Zen4 V-cache could be launched by the end of 2022?

https://twitter.com/x/status/1539171315940204544

https://twitter.com/x/status/1539179107090853888

Yes, AMD said themselves in FAD.

 

[moinmoin]

Since they launched both the 5800X and 5800X3D @ $549, the difference doesn't have to be massive. Could be 7800X @ $499 and 7800X3D @ $599.

I don't expect AMD to reduce prices gen over gen. $100 premium for X3D could happen, though.

 

[eek2121]

With clocks shooting up for Ryzen, and core counts going up for Genoa, I'm rather curious as to which path Zen 4 Threadripper will take. Of course, we won't find out for quite a while. Part of me hopes they release SKUs targeting both core density as well as clocks. A 64 core TR with a peak boost of 5.5 ghz would be a nice little chip to put in my machine, even if it costs 5 figures for the build. 🤣

 

[CakeMonster]

I can't see how it would make sense to launch V-Cache models just a few months after the regular launch. Especially not if its a full lineup, that would make more sense for 2023. But even just one model like 7800X3D would look strange in a brand new lineup.

 

[lobz]

I can't see how it would make sense to launch V-Cache models just a few months after the regular launch. Especially not if its a full lineup, that would make more sense for 2023. But even just one model like 7800X3D would look strange in a brand new lineup.

Marketing and the often underrated significance of virtual bragging rights can always override common sense.

 

[Saylick]

I can't see how it would make sense to launch V-Cache models just a few months after the regular launch. Especially not if its a full lineup, that would make more sense for 2023. But even just one model like 7800X3D would look strange in a brand new lineup.

Why would it be weird to launch V-cache so early? At the end of the day, it's just another feature that can be used to differentiate itself from the rest of the lineup, much like how Intel offered HT models in their i7 CPUs for $100 more than the non-HT versions in their i5 series back in the day. For people who want V-cache and can take advantage of it, they're likely to wait to buy the V-cache variant anyways, so you might as well make it readily available to them ASAP. For those who don't need V-cache, there will always be the vanilla versions.

With the way AMD are rolling out vanilla Zen 4, V-cache Zen 4, and high core count Zen 4, they are trying to cast a wide net so that they offer something for everyone.

 

[biostud]

I don't expect AMD to reduce prices gen over gen. $100 premium for X3D could happen, though.

Maybe not, but price wise they have to compete with 5800X3D and DDR4 though.

 

[randomhero]

I can't see how it would make sense to launch V-Cache models just a few months after the regular launch. Especially not if its a full lineup, that would make more sense for 2023. But even just one model like 7800X3D would look strange in a brand new lineup.

I believe Zen5 will come soon(by soon I mean Q12023), so why wait if it is ready?
There is certainly market that will gobble them up.
Edit: Brain fart, meant Q12024

 

[BTRY B 529th FA BN]

So a rumor is going around saying AM4 might get Zen4 cores on AM4 socket with Zen3 IoD that will utilize DDR4

AMD Ryzen 7000 "Zen 4" Launch Date and Lineup Revealed, Spectacular AM4 Rumor Surfaces

15th September, 2022, is when AMD will debut its Ryzen 7000 "Zen 4" desktop processors. The launch strategy of these chips looks similar to that of the Ryzen 5000 series. The company is preparing a lean launch lineup with just four SKUs—the Ryzen 9 7950X, the Ryzen 9 7900X, Ryzen 7 7800X, and...

www.techpowerup.com