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!

 

[igor_kavinski]

https://twitter.com/x/status/1561971310443794432

If true, 7600X will beat 5800X without even needing to boost. With both hands tied behind its back and kicking the crap out of 5800X. Ouch!

 

[inf64]

https://twitter.com/x/status/1561971310443794432

If true, 7600X will beat 5800X without even needing to boost. With both hands tied behind its back and kicking the crap out of 5800X. Ouch!

We still have no benchmarks though, I wonder why these "leakers" only post pictures of the CPUs and nothing else.

 

[igor_kavinski]

We still have no benchmarks though, I wonder why these "leakers" only post pictures of the CPUs and nothing else.

Presumably they have a lot more to lose than Intel leakers. AMD's NDAs must be scary

 

[inf64]

Presumably they have a lot more to lose than Intel leakers. AMD's NDAs must be scary

Well they have already breached the NDA by posting those pictures...

 

[jamescox]

I remember 6 channels. Nothing official though...

How would you do 6 channels if it has 4 numa nodes per IO die? There are a lot of possibilities with the IO die having ridiculous connectivity, but 8 chiplets should mean 4 quadrants which should have 4, 8, or 12 memory channels. Although, a DDR5 channel is 32-bit, so there may be granularity there to have 3 x 32-bit rather than 6 x 32-bit per quadrant. I am not sure if that is possible though? Haven’t read much on DDR5. It isn’t half the cores if it allows 8 chiplets, so going to 8 channel seems to make sense.

 

[Thibsie]

How would you do 6 channels if it has 4 numa nodes per IO die? There are a lot of possibilities with the IO die having ridiculous connectivity, but 8 chiplets should mean 4 quadrants which should have 4, 8, or 12 memory channels. Although, a DDR5 channel is 32-bit, so there may be granularity there to have 3 x 32-bit rather than 6 x 32-bit per quadrant. I am not sure if that is possible though? Haven’t read much on DDR5. It isn’t half the cores if it allows 8 chiplets, so going to 8 channel seems to make sense.

Hey, that's what I remember I read (no, won't say where). I never said it made a lot of sense to me.

Correction : remembered 6 but it is not.
Checked : Four of'em is correct though. Makes more sense, indeed.

 

[Tuna-Fish]

Thanks. Did they not say that 3000 will be the sweetspot?

Yes, but AMD also said that 2000MHz will be the sweetspot for Zen3. (edit: no they didn't) A lot of Zen3 chips won't run stable above 1800MHz.

I think it's a good idea to wait until launch and independent reviews before going and purchasing memory. It's almost certain that 1:1 at a lower ram speed will beat 2:1 at a higher one in everything except extremely synthetic bandwidth tests, so the best ram you can buy for Zen4 is something that runs the highest clock that the cpu can do at 1:1 and the lowest latencies.

 

[DrMrLordX]

Its a crap test IMO.

Tbh Geekbench has always felt like a better cellphone/light desktop benchmark than something you'd run on a high-end consumer CPU, much less a server/workstation CPU. Even my old 3900X averages ~80W power consumption running the MT test, with the peak being only 125W (out of 142W possible, which is something even Cinebench easily achieves). The CPU is clearly leaving execution resources on the table.

I will have to rerun it on Raphael-X later once I get one to see how bad it is on AM5.

 

[Timmah!]

Yes, but AMD also said that 2000MHz will be the sweetspot for Zen3. A lot of Zen3 chips won't run stable above 1800MHz.

I think it's a good idea to wait until launch and independent reviews before going and purchasing memory. It's almost certain that 1:1 at a lower ram speed will beat 2:1 at a higher one in everything except extremely synthetic bandwidth tests, so the best ram you can buy for Zen4 is something that runs the highest clock that the cpu can do at 1:1 and the lowest latencies.

I see. Did not know that, not owning Ryzen CPU so far, thanks.

 

[Thibsie]

Indeed. Sweet spot doesn't mean everyone will get to it.
Zen4 might be different, we'll see.

 

[carrotmania]

Yes, but AMD also said that 2000MHz will be the sweetspot for Zen3. A lot of Zen3 chips won't run stable above 1800MHz.

No, they absolutely did not.

What is the fastest possible RAM you can run with 1:1 fclk?

DDR4-4000. 2000MHz fclk is to the Ryzen 5000 Series what 1900MHz fclk was to the Ryzen 3000 Series. Many samples can do it, but not all. In upcoming AGESAs, we will be implementing additional tuning that will make reaching 2000MHz easier. This is not a guarantee. There is no “safe bet” that your CPU will reach this frequecy. But don’t be surprised to see a bunch of people bringing that magical 2000MHz fclk home with DDR4-4000.

Okay, so what’s the best price/performance?
DDR4-3600 continues to be a “sweet spot.”

AMD Robert Hallock promises Ryzen 5000 undervolting with new functionality - VideoCardz.com

AMD Director of Technical Marketing – Robert Hallock – provided more details on Ryzen 5000 series compatibility, upcoming features, and upgrade opportunities for users with AMD 500 series motherboards. AMD lifts the curtain on Ryzen 5000 series Yesterday AMD released its Ryzen 5000 series...

videocardz.com

1800FCLK on both Zen2 and Zen3 is the sweet spot, but whereas Zen2 became unstable at 1900mhz, Zen3 could do another 100mhz extra ON A GOOD SAMPLE.

 

[nicalandia]

Genoa Wrecking all kinds of havoc at Geekbench5 Top Multi Score

Top Multi-Core Geekbench 5 CPU Results

Geekbench 5 Results - Geekbench Browser

Over 96,000 points in Geekbench is just crazy even knowing that Geekbench does not scale linearly.

 

[biostud]

Tbh Geekbench has always felt like a better cellphone/light desktop benchmark than something you'd run on a high-end consumer CPU, much less a server/workstation CPU. Even my old 3900X averages ~80W power consumption running the MT test, with the peak being only 125W (out of 142W possible, which is something even Cinebench easily achieves). The CPU is clearly leaving execution resources on the table.

I will have to rerun it on Raphael-X later once I get one to see how bad it is on AM5.

Oh noes, how will I know if a 96 core Epyc system will be my next system???

 

[Asterox]

Genoa Wrecking all kinds of havoc at Geekbench5 Top Multi Score

Top Multi-Core Geekbench 5 CPU Results

Geekbench 5 Results - Geekbench Browser

Over 96,000 points in Geekbench is just crazy even knowing that Geekbench does not scale linearly.

Spectacle, but in real world what can you do with Geekbench.

A beauty for the eyes, and the software used.

- Cinema4d/Octane
- After Effects
- Premiere Pro

 

[nicalandia]

So far these have been the only 32C/32T Zen4 CPUs registered on GB5,

I believe that the ones that show 32.0 MB x 4 are Genoa and the one with 16.0 MB x 4 are Bergamo.

AuthenticAMD Family 25 Model 16(Stepping 0 and 1) are Genoa. AuthenticAMD Family 25 Model 17 are Bergamo

 

[Mopetar]

We still have no benchmarks though, I wonder why these "leakers" only post pictures of the CPUs and nothing else.

Could be a lot of different reasons. The person taking the picture may not be someone who can run any tests on it, the firmware they have might be gimped in some way to make benchmarks meaningless or impossible, or any number of other reasons such as not wanting to draw too much heat from AMD. No one is going to care overly about some pictures of the die that build hype and act like free marketing in a way.

 

[inf64]

Asus NA channel has a live preview of AM5 TUF GAMING X670E-PLUS WIFI for AMD Ryzen 7000 Series Processors:

 

[CakeMonster]

I have the X570-E and very happy with that, so the specs of this looks really good building on that. I'm about half way through, he said PCIE5 on both GPU and NVME simultaneously, I hope that was correct which would be good for longevity. Obviously the 3 PCIE5 NVME will probably not be able to run simultaneously, or at least not with PCIE5 GPU. I suspect we're in for a price shock on the whole X670(E) lineup though.

 

[inf64]

I have the X570-E and very happy with that, so the specs of this looks really. I'm about half way through, he said PCIE5 on both GPU and NVME simultaneously, I hope that was correct which would be good for longevity. Obviously the 3 PCIE5 NVME will probably not be able to run simultaneously, or at least not with PCIE GPU. I suspect we're in for a price shock on the whole X670(E) lineup though.

The initial pricing will probably be out of whack for a while. But if Ryzen 7000 performs great, then the demand will be very high, at least initially.
I'm also very happy with my Asus Prime X570-P, it's a nice board (nothing fancy). I had zero hiccups or failures, even when OCing the lowly 1600AF.

 

[yuri69]

I believe that the ones that show 32.0 MB x 4 are Genoa and the one with 16.0 MB x 4 are Bergamo.

AuthenticAMD Family 25 Model 16(Stepping 0 and 1) are Genoa. AuthenticAMD Family 25 Model 17 are Bergamo

This is odd. The Genoa initial model is Family 25 Model 16 and the last one is Model 31. However, Bergamo should be Family 25 Model 256 and more.

I'd believe that tweet since Bergamo is rumored to be a completely different CCD & packaging (?).

 

[Tuna-Fish]

No, they absolutely did not.

You are correct. This was some mandela effect happening or something, I could have sworn I personally heard AMD talk of 2000MHz in some event, but I looked it up and yeah they only mentioned it as upper limit that only the best examples might reach.

I am now more optimistic about DRAM frequencies for Zen4.

 

[Hans Gruber]

No, they absolutely did not.

What is the fastest possible RAM you can run with 1:1 fclk?

DDR4-4000. 2000MHz fclk is to the Ryzen 5000 Series what 1900MHz fclk was to the Ryzen 3000 Series. Many samples can do it, but not all. In upcoming AGESAs, we will be implementing additional tuning that will make reaching 2000MHz easier. This is not a guarantee. There is no “safe bet” that your CPU will reach this frequecy. But don’t be surprised to see a bunch of people bringing that magical 2000MHz fclk home with DDR4-4000.

Okay, so what’s the best price/performance?
DDR4-3600 continues to be a “sweet spot.”

AMD Robert Hallock promises Ryzen 5000 undervolting with new functionality - VideoCardz.com

AMD Director of Technical Marketing – Robert Hallock – provided more details on Ryzen 5000 series compatibility, upcoming features, and upgrade opportunities for users with AMD 500 series motherboards. AMD lifts the curtain on Ryzen 5000 series Yesterday AMD released its Ryzen 5000 series...

videocardz.com

1800FCLK on both Zen2 and Zen3 is the sweet spot, but whereas Zen2 became unstable at 1900mhz, Zen3 could do another 100mhz extra ON A GOOD SAMPLE.

I endorse MSI and Asus motherboards. With that said any DDR4 3600mhz binned memory can run 3800mhz without issues on B350/x370 all the way up to B550/x570.

I have e-die sticks that run 3800mhz without any issue and they are binned at 3000mhz. I have Hynix CJR sticks binned @3600mhz that run a Zen 2 B350 @ 3800mhz without any issue. 16-20-20 on those.

3600mhz was the sweet spot only because of the AGESA bios issues beyond 3600mhz. AMD fixed on the latest AGESA release early in 2022.

I have heard rumors that 2000mhz on Zen 2 and Zen 3 would be possible without the memory divider issue introducing a lot of latency above 3800mhz. Can someone clarify this?

 

[jamescox]

This is odd. The Genoa initial model is Family 25 Model 16 and the last one is Model 31. However, Bergamo should be Family 25 Model 256 and more.

I'd believe that tweet since Bergamo is rumored to be a completely different CCD & packaging (?).

It is supposed to be a 16 core CCD, probably with two 8 core CCX per die. Packaging should essentially be the same as Genoa; likely with the same IO die. It surprises me that Bergamo samples seem to be around already. I thought it would be a while after Genoa, but perhaps it will be earlier than expected. AMD seems to be able to control the information very well. Intel has been leaking all kinds of things, but that is deliberate to try and keep in the news.

 

[Joe NYC]

Zen 4 3D V-Cache blows the doors off

 

[Vope45]

30% faster than standard zen 4 holly molly. I hope this is true.