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!

 

[inf64]

@inf64
graph made by AMD fanboy as many here still promote and FUD intel for their own reason, there is <1% chance 16/32 7950x is faster then 24/32 13900k

I'm not sure who are you referring to, but I don't fud intel or "promote" AMD. This is just my hobby, and I did it for fun. Let's circle back to this post after Raptor Lake launches so we will see how much I failed in my estimate.

Why don't you enlighten us and let us know how each will perform in your expert opinion?

 

[moinmoin]

@inf64
graph made by AMD fanboy as many here still promote and FUD intel for their own reason, there is <1% chance 16/32 7950x is faster then 24/32 13900k

You are in the Zen 4 thread. Furthermore E-cores are commonly equaled to 1/2 P-cores or one P-core thread, so 8 P-cores + 16 E-cores could be summarized as 16 big cores. What makes you so sure there is "<1% chance"?

 

[Schmide]

I have an excel chart with all the mental gymnastics, so if anyone is interested here it is: https://docs.google.com/spreadsheet...ouid=117779538430217342497&rtpof=true&sd=true

My only gripe is corrected should be replaced with normalized to. It took me a few secs to get it.

 

[inf64]

My only gripe is corrected should be replaced with normalized to. It took me a few secs to get it.

That's a valid point, I changed it. Thank you!

 

[Markfw]

You are in the Zen 4 thread. Furthermore E-cores are commonly equaled to 1/2 P-cores or one P-core thread, so 8 P-cores + 16 E-cores could be summarized as 16 big cores. What makes you so sure there is "<1% chance"?

Exactly ! 2 E-cores are barely equal to one P-core. In many cases they fare even worse. I have one, I know.(mine is 8+4, 12700F)

 

[VirtualLarry]

This forum gets mentioned by Paul's Hardware.

(Edit: for a snapshot of 7000-series Zen4 chip)

 

[Timmah!]

This forum gets mentioned by Paul's Hardware.

Interwebz fame is upon all of us, at last. Rejoice, wherever you are!

 

[igor_kavinski]

Wow. Didn't think I would learn something from that. The review sample was leaked by Lenovo and Janet Jackson's song Rhythm Nation caused 5400 rpm hdds to crash. WOAH DUDE!

 

[DrMrLordX]

graph made by AMD fanboy as many here still promote and FUD intel for their own reason, there is <1% chance 16/32 7950x is faster then 24/32 13900k

Tsk tsk "fanboy" was a bannable awhile back. Be careful.





It still is against the rules


esquared
Anandtech Forum Director

 

[cellarnoise]

There is a difference between single / (really 2 or 3 threaded programs - as what with a modern o.s. is single?), threaded programs that have parts or all that are multi-threaded, and running programs that are almost entirely different threads doing the same general compute, but in separate threads.

D.C. is often a good example and like virtual server use - number of cores fully powered is in constant flux. Often this is really just x# of single threaded or more - puters within the same package / motherboard / and stuff needing shared cpu time.

Multi-threaded examples seem to be more limiting anymore, as cpus have ever more cores and threads. Really need better independent times x threaded benchmarks using parts of a multi-cpu / threaded examples.

I know this is the trend in benchmarking large core cpu's but not fully the norm yet?

 

[reb0rn]

You are in the Zen 4 thread. Furthermore E-cores are commonly equaled to 1/2 P-cores or one P-core thread, so 8 P-cores + 16 E-cores could be summarized as 16 big cores. What makes you so sure there is "<1% chance"?

I would not agree for that math, as E core use a lot less power so in multithread work it can push the limits while the zen and P cores will be limited by the power more so
also I almost never see point in thread in task I need

 

[FangBLade]

@inf64
graph made by AMD fanboy as many here still promote and FUD intel for their own reason, there is <1% chance 16/32 7950x is faster then 24/32 13900k






Use of word fanboy in the tech forums is prohibited.


esquared
Anandtech Forum Director

Your first comment here, and yet so offensive. Look at yourself first before commenting on others, have a nice day.

 

[TESKATLIPOKA]

I would not agree for that math, as E core use a lot less power so in multithread work it can push the limits while the zen and P cores will be limited by the power more so
also I almost never see point in thread in task I need

E-core is not more power efficient than Zen core, It just consumes less.
Power consumption in MT should look like this.
8 E-cores at 3.7-3.9GHz consume ~50W or 6.125W per core.
16 Zen3 ~3.8-4GHz consume and 130-142W or 8.125-8.875W per core.
Difference in power consumption is 33-45%, but performance per core is still higher.
Even compared to Zen 4, the E-cores shouldn't be more power efficient.

 

[Zucker2k]

I'm not sure who are you referring to, but I don't fud intel or "promote" AMD. This is just my hobby, and I did it for fun. Let's circle back to this post after Raptor Lake launches so we will see how much I failed in my estimate.

Why don't you enlighten us and let us know how each will perform in your expert opinion?

I don't have a problem with this estimate of yours. In fact, I'm pleasantly surprised by how much your position has changed on your initial projections in favour of RPL. We've gone from Zen 4 "destroying" RPL to what is now a virtual tie. How many IPC revisions of your IPC calculations did it take to get here because I lost count, frankly.

Zen 4 is definitely going to be competitive for the mere fact that it's going to be able to maintain high clocks in multithreaded scenarios, much better than Raptor Cove. The dark horse most people have failed to take into consideration, including you, are the e-cores. The simple math is this:
In multithreaded workloads, and assuming the 8 p-cores of the Raptor Cove cores can neutralize 8 Zen 4 cores, then the rest boils down to 8 Zen 4 cores vs 16 (cache) upgraded e-cores. If Zen 4 holds 4.5Ghz, all-core, a reasonable and modest proposition, the e-cores merely have to run 3Ghz, which amounts to 6Ghz x8 in processing power.

So, I believe all this really boils down to how high Zen 4 cores can hold clocks vs how high the e-cores can hold clocks, with the headroom going to the e-cores, in my opinion, just because these upgraded e-cores are going to be reasonably faster than skylake cores. Shouldn't be long before we find out.

Edit: Let me add that I haven't accounted for a lack of HT on the e-cores. However, the e-cores have quite a lot of headroom so things should be interesting.

 

[TESKATLIPOKA]

I don't have a problem with this estimate of yours. In fact, I'm pleasantly surprised by how much your position has changed on your initial projections in favour of RPL. We've gone from Zen 4 "destroying" RPL to what is now a virtual tie. How many IPC revisions of your IPC calculations did it take to get here because I lost count, frankly.

Zen 4 is definitely going to be competitive for the mere fact that it's going to be able to high clocks in multithreaded scenarios, much better than Raptor Cove. The dark horse most people have failed to take into consideration, including you, are the e-cores. The simple math is this:
In multithreaded workloads, and assuming the 8 p-cores of the Raptor Cove cores can neutralize 8 Zen 4 cores, then the rest boils down to 8 Zen 4 has cores vs 16 (cache) upgraded e-cores. If Zen 4 holds 4.5Ghz, all-core, a reasonable and modest proposition, the e-cores merely have to run 3Ghz, which amounts to 6Ghz x8 in processing power.

So, I believe all this really boils down to how high Zen 4 cores can hold clocks vs how high the e-cores can hold clocks, with the headroom going to the e-cores, in my opinion, just because these upgraded e-cores are going to be reasonably faster than skylake cores. Shouldn't be long before we find out.

If we talk about 32 threads then It will be tie as you say, but what about 16 threads? That's the worst possible scenario I can think of.
8P cores vs 8 Zen4 cores will neutralize each other.
Then you have 8 E cores vs 8 Zen 4 cores, they will give you maybe 40% of Zen 4 performance.
The end result will be 30-35% more performance for Zen4.

Markfw: You have both ADL i7-12700F and Zen 3 5950X, right? Can you please test your Zen 3 in CB R23 from 1T->16T, and ADL 1T->12T and upload the results?

If anyone has i9-12900, then please test It in CBR23 from 1 thread up to 16 threads and upload the results.
Then we can compare how ADL compares to Zen3 when only cores are used.
I think everyone would love to see the results. Many thanks.

P.S. Actually It can be up to max threads, so we can see how much SMT(HT) provides. Thanks

 

[Timorous]

There was that rumour a while ago that AMD could make some Zen 4 parts available on AM4 as budget options, personally I don't see it but what if AMD made Zen 3 available on AM5 for budget options.

In theory AMD can just put Zen 3 CCDs on the package rather than Zen 4 CCDs and make a budget AM5 offering to get people onto the platform and give them options for drop in CPU upgrades later like when Zen 5 or Zen 6 comes out.

There is also the rumour that AMD are planning on dropping the 5800X3D price due to how good Zen 4 is at gaming, if that is true then why not make a 5600X3D and 5800X3D AM5 option to fill in the budget gaming end of the market. It would probably offer gaming parity with top tier ADL and mid/high tier RKL but it could come in a lot lot cheaper while getting people to buy into a new platform at the same time making future drop in CPU upgrades more likely even if Intel can stay competitive.

 

[Zucker2k]

If we talk about 32 threads then It will be tie as you say, but what about 16 threads? That's the worst possible scenario I can think of.
8P cores vs 8 Zen4 cores will neutralize each other.
Then you have 8 E cores vs 8 Zen 4 cores, they will give you maybe 40% of Zen 4 performance.
The end result will be 30-35% more performance for Zen4.

Markfw: You have both ADL i7-12700F and Zen 3 5950X, right? Can you please test your Zen 3 in CB R23 from 1T->16T, and ADL 1T->12T and upload the results?

If anyone has i9-12900, then please test It in CBR23 from 1 thread up to 16 threads and upload the results.
Then we can compare how ADL compares to Zen3 when only cores are used.
I think everyone would love to see the results. Many thanks.

P.S. Actually It can be up to max threads, so we can see how much SMT(HT) provides. Thanks

I think the comparison you're looking for is a 12900k vs a 5900x. Zen 4 will have a higher IPC and higher clocks.

 

[coercitiv]

If Zen 4 holds 4.5Ghz, all-core, a reasonable and modest proposition

4.5Ghz @ 230W is something that 5950X can do today in CB / Blender, expect Zen 4 to clock higher than this given the new node and the speed oriented design.

 

[jpiniero]

There was that rumour a while ago that AMD could make some Zen 4 parts available on AM4 as budget options, personally I don't see it but what if AMD made Zen 3 available on AM5 for budget options.

I think AMD might hold off on any budget options because of DDR5 prices... but I suspect the budget option will be Rembrandt with the IGP disabled.

 

[Tuna-Fish]

In theory AMD can just put Zen 3 CCDs on the package rather than Zen 4 CCDs and make a budget AM5 offering to get people onto the platform

This really doesn't work, because the major cost in the platform is DDR5 and the motherboards. Having a CPU cheaper than 7600X would really be a blib in the total platform cost.

 

[TESKATLIPOKA]

I think the comparison you're looking for is a 12900k vs a 5900x. Zen 4 will have a higher IPC and higher clocks.

What I am looking for is 12900k vs 5950x, but Markfw only has 12700.
I am interested in CB R23 score at different number of threads, because in reviews you either have 1T or Max number of threads and nothing between.



.

 

[uzzi38]

I don't have a problem with this estimate of yours. In fact, I'm pleasantly surprised by how much your position has changed on your initial projections in favour of RPL. We've gone from Zen 4 "destroying" RPL to what is now a virtual tie. How many IPC revisions of your IPC calculations did it take to get here because I lost count, frankly.

Zen 4 is definitely going to be competitive for the mere fact that it's going to be able to maintain high clocks in multithreaded scenarios, much better than Raptor Cove. The dark horse most people have failed to take into consideration, including you, are the e-cores. The simple math is this:
In multithreaded workloads, and assuming the 8 p-cores of the Raptor Cove cores can neutralize 8 Zen 4 cores, then the rest boils down to 8 Zen 4 cores vs 16 (cache) upgraded e-cores. If Zen 4 holds 4.5Ghz, all-core, a reasonable and modest proposition, the e-cores merely have to run 3Ghz, which amounts to 6Ghz x8 in processing power.

So, I believe all this really boils down to how high Zen 4 cores can hold clocks vs how high the e-cores can hold clocks, with the headroom going to the e-cores, in my opinion, just because these upgraded e-cores are going to be reasonably faster than skylake cores. Shouldn't be long before we find out.

Edit: Let me add that I haven't accounted for a lack of HT on the e-cores. However, the e-cores have quite a lot of headroom so things should be interesting.

I can tell you now that 4.5GHz all core is going to be extremely easy to achieve.

 

[Abwx]

If Zen 4 holds 4.5Ghz, all-core, a reasonable and modest proposition, the e-cores merely have to run 3Ghz, which amounts to 6Ghz x8 in processing power.

At equal frequency 1 P core is equivalent to 1.8 e-core and Zen 4 will be on the same ballpark as GC, so that require the 16 e-cores to run at 10% lower frequency than 8 P cores to get the same throughput.

In your exemple the 16 e-cores should be clocked at 4.05GHz to match 8 Zen 4 cores@4.5GHz.

 

[Asterox]

I can tell you now that 4.5GHz all core is going to be extremely easy to achieve.

For example, stock R5 5600X hits 4.5 on all cores or 12 threads boost.

AMD Ryzen 5 5600X Review

Six Zen 3 cores beating eight Zen 2 Cores? That's exactly what's happening with the Ryzen 5 5600X. AMD's massive IPC gain helped it overcome a two-core deficitm, even in productivity tests. The Ryzen 5 5600X redefines what you really need for a high-end gaming PC.

www.techpowerup.com

R5 7600X, hm probably minimal 5-5.1ghz all core boost no problem.

 

[TESKATLIPOKA]

I can tell you now that 4.5GHz all core is going to be extremely easy to achieve.

Everyone is discussing only desktop, but In my opinion mobile is where AMD will shine.
Both with Phoenix(strong IGP) and with Raphael(16C).
Intel won't give you a CPU with more than 96EU so Phoenix will win by default.
It's likely Intel will release 8P16E mobile CPU, but power will be a problem as is with ADL and AMD will double the number of cores.