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!

 

[Markfw]

@eek2121
Well, you need to take performance into account as well. If the AVX test is much faster even downclocked, then total energy consumption for a given workload very likely is smaller. The main reason for the behaviour should be that the occupancy of execution units is much higher with AVX.

I know nobody else wants to hear about DC performance (max utilization of all cores 24/7) but if you really want to know, check the BOINC log files in regular and avx-512 tasks. The 7950x in both categories just blows away the 13900k. In points, even the 64 core EPYCs with the current project. Now if I had a 7773x maybe that would all change, but I don't have the money for that. But it does beat the 7763 Milans I have.

 

[Tuna-Fish]

Dark red square is System RAM

And the latency gets worse with a larger working set at the end because the L2 DTLB runs out. 8x PTE coalescing * 3072 pages * 4096kB page size = ~100MB.

 

[mmaenpaa]

@Det0x
Your numbers are a bit better as the ones from CnC. But they again show quite some lack of bandwidth compared to RTL.

@mmaenpaa
So you have this shiny new 7700X. Might you give the Micro Bench posted by @Det0x a try? 😉 I am really interested how a single CCD ZEN4 might turn out bandwidth wise.

System is at Asrock defaults except EXPO @6000. Windows 11 Pro latest as of today. It seems I need to upgarde bios to 1.11.AS03 (beta) which enchances S3 mode (=corp speak, it may now work 🤣)

 

[Det0x]

Interesting. Now that I have my new system built, I've been doing some Aida64 runs and one thing I noticed is that my memory latency scores are higher than what I thought they would be using DDR5 7400. The lowest latency score I've had so far is 61.7ns. This version I am using is a more recent build so they cannot truly be compared, but I definitely think that AMD has the advantage in memory latency due to being able to use Gear 1 with DDR5. If whoever did that run with CL30 DDR5 6000 memory, he would have gotten into the 5xns range for sure.

But Zen 4's read, write and copy scores leave much to be desired to be honest. Not sure what's up with that.

This is how aida64 looks for my maxed out 7950x
(dont know where nicalandia got that other aida64 screenshot, but it's certainly not my system)

System is at Asrock defaults except EXPO @6000. Windows 11 Pro latest as of today. It seems I need to upgarde bios to 1.11.AS03 (beta) which enchances S3 mode (=corp speak, it may now work 🤣)

View attachment 70711View attachment 70712View attachment 70713

Wow thats very low latency, have to give single CCD a try when i get home later today

That is 58.5ns @ ~65mb (not max value that's ~1gigabyte)

 

[Carfax83]

This is how aida64 looks for my maxed out 7950x
(dont know where nicalandia got that other aida64 screenshot, but it's certainly not my system)
View attachment 70714

Now that is definitely some tweaked memory! What voltage are you using to get the timings so low? And yeah, that latency is extremely low to the point that I doubt anything other than a massively tweaked and overclocked Raptor Lake CPU could match it.

BTW, could you also show your cache benchmarks, especially the L3 ones.

Here's one of mine at 7600mhz:

 

[Det0x]

Now that is definitely some tweaked memory! What voltage are you using to get the timings so low? And yeah, that latency is extremely low to the point that I doubt anything other than a massively tweaked and overclocked Raptor Lake CPU could match it.

BTW, could you also show your cache benchmarks, especially the L3 ones.

Here's one of mine at 7600mhz:

Here is one of my older screenshots
6200MT/s 2200mhz FCLK.
(my new settings is 6400MT/s 2133mhz FCLK)

Running 1.53 vdimm with custom cooling on memory

Btw we have a competition thread going over at overclock.net with this new memory benchmark:

New Memory Bandwidth & Latency Competition

I will post my 5800x3d and 7950x with single and dual CCD numbers in that benchmarkthread when i get home from work later today
Would be interesting if you guys here could share your numbers over there also.
I'm especially keen on seeing bandwidth numbers from one of @Markfw's epic systems if possible.. Currently there is a dual socket Xeon laying waste in the bandwidth department over there

 

[Carfax83]

Here is one of my older screenshots
6200MT/s 2200mhz FCLK.
(my new settings is 6400MT/s 2133mhz FCLK)
View attachment 70715
Running 1.53 vdimm with custom cooling on memory

You have to have a lot of knowledge to tweak at such a high level so my hat is off to you. And AMD deserves credit for how brutally fast their L3 cache is with Zen 4.

My own score is somewhat depressed because I have mine underclocked and undervolted. I'm done with balls to the wall overclocking and tweaking though. No longer have the time or patience. I just want things nice and easy these days

Btw we have a competition thread going over at overclock.net with this new memory benchmark:

New Memory Bandwidth & Latency Competition

https://www.overclock.net/threads/new-memory-bandwidth-latency-competition.1801933/#post-29068256

Nothing but hardcore dudes over at Overclock.net to be sure. I used to have a membership over there a long time ago but if my membership is still valid I don't have a clue what my password would have been.

 

[DisEnchantment]

I'm especially keen on seeing bandwidth numbers from one of @Markfw's epic systems if possible

Time for @Markfw to step up to dual socket Genoa X and dual socket SH5 APUs(if they are at all dual socket). SH5 APUs will have V Cache + HBM.
These big boys are only for folks going for the ultimate epeen with the most demanding loads.

 

[BorisTheBlade82]

System is at Asrock defaults except EXPO @6000. Windows 11 Pro latest as of today. It seems I need to upgarde bios to 1.11.AS03 (beta) which enchances S3 mode (=corp speak, it may now work 🤣)

View attachment 70711View attachment 70712View attachment 70713

Thanks a lot. That pretty much confirms that single CCD SKUs are severely bandwidth limited through IFoP (not that this would be a big concern for Games and most other workloads).
Could you do me a favour and test the write bandwidth in isolation as well?

 

[nicalandia]

I will post my 5800x3d and 7950x with single and dual CCD numbers in that benchmarkthread when i get home from work later today

It would be interesting to know the latency numbers for the 5800X3D and compare it with 5800X and see the latency jump on the 5800X when it runs out of 32MiB L3 and jumps to System RAM while the 5800X3D keeps those low until it runs out of L3 at the 96MiB mark

 

[Markfw]

Here is one of my older screenshots
6200MT/s 2200mhz FCLK.
(my new settings is 6400MT/s 2133mhz FCLK)
View attachment 70715
Running 1.53 vdimm with custom cooling on memory

Btw we have a competition thread going over at overclock.net with this new memory benchmark:

New Memory Bandwidth & Latency Competition

I will post my 5800x3d and 7950x with single and dual CCD numbers in that benchmarkthread when i get home from work later today
Would be interesting if you guys here could share your numbers over there also.
I'm especially keen on seeing bandwidth numbers from one of @Markfw's epic systems if possible.. Currently there is a dual socket Xeon laying waste in the bandwidth department over there

They run linux. They will not even start to load win 10 with 256 threads. Do you have a suggestion of what to run ? I can maybe later today. The dual Milan system would probably be the fastest.

 

[Det0x]

Wow thats very low latency, have to give single CCD a try when i get home later today

lol you just hadn't scrolled all the way down..
That was 58.5ns @ ~65mb (not max value which is ~1gigabyte)

7950x single ccd and 5800x3d number coming shortly

 

[Det0x]

5800x3d @ 1800:3600 XMP timings

Latency = 78.82ns


Dram read bandwidth = 37.89 GB/s


7950x single CCD (8cores only) "6200MT/s" 2133 FCLK

Latency = 71.22ns


Dram read bandwidth = 67.70 GB/s


As a side note, it looks like the 5800x3d gets much of its gaming prowess from around the 100mb datasize..
Most games must use that data size.. (?)

V-cache difference is showing from 65mb upto 262mb
(5800x3d vs my "8core" Zen4 below)

Latency:
vs

Bandwidth:
vs

 

[eek2121]

My largely untweaked 7950x (lowered tRFC a small amount, UCLK=MCLK, that is it):

Spoiler

 

[nicalandia]

5800x3d @ 1800:3600 XMP timings


Main v-cache difference is from 65mb upto 262mb

Thanks for sharing those numbers, but I was more interested on the L3 access latency numbers for Zen3X3D vs Zen4 and where most of the gaming gains come from.

Here are the Results Highlighted.

5800X3D


One CCD 7950X




Look at the Latency numbers at 1024 KB mark. Thanks to Zen4 larger L2$ it's just 3.78ns that right there is a huge help on games and general purpose apps(Zen3X3D is 10.75ns)

No lets move to the 98304 KB On Zen4 it's 52.28ns and on Zen3X3D is 26.22ns That has a huge impact on many games(not all of them of course)

 

[nicalandia]

Less than 10 minutes before AMD Releases The Kraken called Genoa.




It will trully be a Blood Bath for poor 8380 Ice Lake Xeons.

 

[nicalandia]

Hey look at this.

AMD 4th Gen EPYC 9004 "Genoa Zen 4" CPUs Launched: Up To 96 Cores, 192 Threads, 384 MB L3 Cache & Crushing All Other Server Chips

AMD EPYC 9004 "Genoa", the 4th Gen Data Center CPU lineup, featuring the latest Zen 4 core architecture has officially made its debut today.

wccftech.com

 

[Markfw]

Makes my dual 7763's look like they are standing still.

 

[mmaenpaa]

Thanks a lot. That pretty much confirms that single CCD SKUs are severely bandwidth limited through IFoP (not that this would be a big concern for Games and most other workloads).
Could you do me a favour and test the write bandwidth in isolation as well?

Did you mean this:

 

[Panino Manino]

Streaming now, should we care?

 

[moinmoin]

Comparison between gens:

 

[nicalandia]

Against SPR-SP will be even more area efficient since Ice lake xeons are 6.5 mm^2, Server grade Golden Cove are about 8 mm^2


This is the AMD EPYC 9654P which is a Single Socket CPU

ASUSTeK COMPUTER INC. RS500A-E12-RS12U - Geekbench

Benchmark results for an ASUSTeK COMPUTER INC. RS500A-E12-RS12U with an AMD EPYC 9654P processor.

browser.geekbench.com

 

[moinmoin]

In my opinion AMD pushes to much the difference between 4th gen Epyc and Intel's available solutions. That constant sales pitch could be summarized with following single slide:

 

[yuri69]

Thanks a lot. That pretty much confirms that single CCD SKUs are severely bandwidth limited through IFoP (not that this would be a big concern for Games and most other workloads).

It's a letdown. Zen 4 client was rumored to use an advanced interconnect. Another point for Zen 5 musthaves.

 

[igor_kavinski]

Makes my dual 7763's look like they are standing still.

It's a long shot but you could write to AMD to donate a Genoa server to you for the noble pursuit of science and medicine. Just show them photos of the servers you already have. I think they would be interested in publicity like that.