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

Page 496 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Vattila

Senior member
Oct 22, 2004
799
1,351
136
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).

Untitled2.png


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! :)
 
Last edited:
  • Like
Reactions: richardllewis_01

Joe NYC

Golden Member
Jun 26, 2021
1,872
2,153
106
It is very reasonable to assume, that they might make it happen (twice the cores on the same area. I also suspect them to use another IO - either full on SoIC or InFO-RDL.

Pure speculation:
I measured the block on the MI300 render that I suspect to be the 16c Zen4c CCD and it nicely fits the area bill (see the Zen5 thread around two weeks back, as ATM I am just too lazy).

My latest theory on the CPU cores of Mi300 is that the graphical representation is probably incorrect and it is 3 Zen4 chiplets of 8 cores each.

It would require the least amount of changes from the existing Zen 4 chiplet.
 
  • Like
Reactions: Vattila

trivik12

Senior member
Jan 26, 2006
250
201
116
Are there any Zen 4 laptops available for sale now anywhere. I dont see anything at amazon or bestbuy online. Forget instore as I dont think I have ever seen even rembrandt laptops in store at Best buy or Costco. I have seen few Zen 2/Zen 3 laptops.
 

trivik12

Senior member
Jan 26, 2006
250
201
116
For some weird reason, AMD is less focused on laptops when it comes to ramp ups. They take eons and availability is also spotty. its lot easier to buy Intel laptops or Macbooks here in US.
 

moinmoin

Diamond Member
Jun 1, 2017
4,926
7,608
136
For some weird reason, AMD is less focused on laptops when it comes to ramp ups. They take eons and availability is also spotty. its lot easier to buy Intel laptops or Macbooks here in US.
I don't think it's weird, laptops are a tight supply chain where everything needs to come together (with AIO and even more so desktop it's much easier to second source and mix 'n' match) and it's obvious Intel both offers the quantity and the leeway to ensure more timely releases by partners. AMD isn't bad considering where it was just a couple years ago, but the focus on few specific partners (ASUS, Lenovo, HP) shows it's not exactly easy to scale that effort wide and up. So most other manufacturers come significantly later, and as Rembrandt showed economic circumstances can lead to that step barely happening.
 

nicalandia

Diamond Member
Jan 10, 2019
3,330
5,281
136
Anyone here remember AMD Athlon FX 4x4?
279c-amd-4x4.gif

I believe AMD had that in mind when making AMD Threadripper Pro 3000 series just in case Intel pulled a Rabbit out of the Hat in the HEDT Segment.

Screenshot_20230315-053449_Gallery.jpg


The TR 5000 series is also capable and likely the Zen4 based TR models as well
 

DrMrLordX

Lifer
Apr 27, 2000
21,570
10,762
136
I don't think it's weird, laptops are a tight supply chain where everything needs to come together (with AIO and even more so desktop it's much easier to second source and mix 'n' match) and it's obvious Intel both offers the quantity and the leeway to ensure more timely releases by partners. AMD isn't bad considering where it was just a couple years ago, but the focus on few specific partners (ASUS, Lenovo, HP) shows it's not exactly easy to scale that effort wide and up. So most other manufacturers come significantly later, and as Rembrandt showed economic circumstances can lead to that step barely happening.

I have suspected that AMD prioritizes bulk corporate buys first, to the detriment of general retail markets.
 
  • Like
Reactions: Tlh97 and Ajay

moinmoin

Diamond Member
Jun 1, 2017
4,926
7,608
136
I have suspected that AMD prioritizes bulk corporate buys first, to the detriment of general retail markets.
I'd claim that's effectively the case anyway. Whether it's by AMD's doing or the manufacturers' doing it just makes perfect business sense to sell in bulk wherever possible.
 

DrMrLordX

Lifer
Apr 27, 2000
21,570
10,762
136
I'd claim that's effectively the case anyway. Whether it's by AMD's doing or the manufacturers' doing it just makes perfect business sense to sell in bulk wherever possible.

True. I still wonder, is there an actual product shortage though, or is that the market AMD prioritizes, thereby limiting how many Dragon Range or Phoenix units they'll ship? If all AMD cares about is the buik buys, all they need to do is gauge demand from the OEMs that ship the finished units, and then feed them accordingly. AMD shouldn't have any trouble getting enough wafers to serve the broader retail market, but under the current circumstances, maybe they're not interested in pushing too much hardware out into general retail space.
 

moinmoin

Diamond Member
Jun 1, 2017
4,926
7,608
136
True. I still wonder, is there an actual product shortage though, or is that the market AMD prioritizes, thereby limiting how many Dragon Range or Phoenix units they'll ship? If all AMD cares about is the buik buys, all they need to do is gauge demand from the OEMs that ship the finished units, and then feed them accordingly. AMD shouldn't have any trouble getting enough wafers to serve the broader retail market, but under the current circumstances, maybe they're not interested in pushing too much hardware out into general retail space.
In my eyes as long as there is more demand than supply there is a shortage. If the output is only sufficient to fulfil the bulk corporate demand then there's shortage of output for the general retail space. That may be on purpose, but I honestly can't imagine continued Rembrandt's scarceness as being on purpose. In any case it would be really interesting to hear what exactly is the limiting factor.
 

Det0x

Golden Member
Sep 11, 2014
1,027
2,953
136
For those interested in where the Zen4 X3D gaming performance really comes from, ive made a post about it here
tldr
  • 7950X3D
  • 6600MT/s tight timings
  • 2166mhz FCLK
Read bandwidth = 91.62 GB/s
1678983438320.png

Default pages = 69.59ns latency
1678983503884.png

Vanilla Zen4 bandwidth--- VS --------- X3D Zen4 bandwidth
1678983314706.png VS 1678983337579.png

Vanilla Zen4 latency-------- VS -----------------X3D Zen4 latency
1678983403320.png VS 1678983414378.png

All this extra gaming performance in Zen4 7950x3d comes from this little difference :cool:
 
Last edited:

biostud

Lifer
Feb 27, 2003
18,171
4,648
136
For those interested where the Zen4 X3D gaming performance really comes from, ive made a post about it here
tldr
  • 7950X3D
  • 6600MT/s tight timings
  • 2166mhz FCLK
Read bandwidth = 91.62 GB/s
View attachment 78252

Default pages = 69.59ns latency
View attachment 78253

Vanilla Zen4 bandwidth--- VS --------- X3D Zen4 bandwidth
View attachment 78248 VS View attachment 78249

Vanilla Zen4 latency-------- VS -----------------X3D Zen4 latency
View attachment 78250 VS View attachment 78251

All this extra gaming performance in Zen4 7950x3d comes from this little difference :cool:
What are the numbers for a 13900k?
 

Joe NYC

Golden Member
Jun 26, 2021
1,872
2,153
106
Not sure if people caught this article about a Ryzen server from Asrock.
AMD Ryzen Server the ASRock Rack 1U4LW-B650/2L2T Review (servethehome.com)

So, I ordered the mobo only for a replacement of a small server in our office. I am going to pair it with a 7700x combo with a free 32GB of DDR5 6000 from Micro Center and an M.2 drive I have sitting around (for now). Should be under $1000.

I may eventually replace the M.2 drive with PCIe 5, when the good ones come out.
 
Last edited:

Joe NYC

Golden Member
Jun 26, 2021
1,872
2,153
106
The board itself is $490 on Newegg. And the memory QVL includes 1 16GB stick of ECC UDIMM. I'll wait :rolleyes:

I think the price (compared to consumer motherboards) has to do with extra networking features, 2x10 GBase-T Broadcom + 1x2.5 Intel + management.

It has all 24 PCIe Gen 5 lanes exposed:
16x PCIe
4x PCIe
4x M.2

The board also takes non-ECC DIMMs, FYI, if you don't want to limit yourself to ECC.
 

nicalandia

Diamond Member
Jan 10, 2019
3,330
5,281
136
There's something off about those results. Why would populating the last 2 channels have such an outsized impact?
Who knows, but HPCG is very bandwith sensitive. And by the way, those are the result with HPCG Compiled with generic compiler, compiled with specific Zen4 flags the performance is about 40% Higher.

Having said that HPCG, HPL, NAMD, ML and many Scientific oriented tasks are best procesed by GPUs

Look at this.pic_disp (2).jpeg

Even with Optimized build HPCG 2S Genoa is slower than a single 3080 Ti GPU, a pair of Xeon 8490H fair far worse of course
 
  • Like
Reactions: Tlh97

lightmanek

Senior member
Feb 19, 2017
387
754
136
Sure, but those results clearly aren't just scaling with the change in bandwidth. There must be something else happening for the 12 channel config. Maybe related to the IO die partitioning?

Yes, I bet it is due to some chiplets having to cross NUMA nodes. Having each die direct access to it's own memory controller must be making the difference. It would be interesting to see this same benchmark on lesser Genoa model with 6 channels and half the chiplets.
 

nicalandia

Diamond Member
Jan 10, 2019
3,330
5,281
136
Yes, I bet it is due to some chiplets having to cross NUMA nodes. Having each die direct access to it's own memory controller must be making the difference. It would be interesting to see this same benchmark on lesser Genoa model with 6 channels and half the chiplets.
This is the result of the 7950X with generic build and with optimized Zen4 build.

hpcg-zen4-768x422 (1).png

The 7950X having only 2 Channels gets a rather low score