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!

 

[Timorous]

I was only referring to fps, when talking about CPU limitations and also only desktop computers.

Why only fps? Seems a bit limiting and is something that should be solved. I think there is a market for a channel to cater to the wider gaming audience and get a lot of views. If some had a solid methodology and dropped all the productivity benchmarks from CPU benchmarks to replace them with tic-rate and turn-time gaming tests I could see it being successful.

 

[nicalandia]

WTF does anyone find this sort of off?



I could of sworn the WR for threadripper was 5.15ghz..

On Geekbench own search option. That's the Fastest Threadripper PRO recorded(O.C) vs Stock 9554 64 Core Genoa

 

[biostud]

Why only fps? Seems a bit limiting and is something that should be solved. I think there is a market for a channel to cater to the wider gaming audience and get a lot of views. If some had a solid methodology and dropped all the productivity benchmarks from CPU benchmarks to replace them with tic-rate and turn-time gaming tests I could see it being successful.

I agree, that it would be a good ide that more CPU reviews focused on games that taxed the CPU in another way, not related to fps. So it is just that my focus was in that limited scope that is fps, which most game benchmarks are benchmarking and then use other software to benchmark other aspects of CPU load. So if there were a couple of popular CPU demanding games, that were easily put into a benchmark suite, which focused on turn time instead of fps, I would be all for it. I'm just not aware of anyone who does that currently.

 

[nicalandia]

Cinebench R23 Score on a 100-000000897-03 (EPYC 9334 32 Cores ES Sample), Don't mind CPU-Z Reading Ryzen(it's a misread)

50,000 Points is really impressive for such single low speed CPU.

 

[Kocicak]

... EPYC 9334 32 Cores ES Sample ... 50,000 Points is really impressive for such single low speed CPU.

50 000 points at what power draw? At 150W that would be very impressive. 50 000 points alone mean nothing.

 

[nicalandia]

50 000 points at what power draw? At 150W that would be very impressive. 50 000 points alone mean nothing.

Stock 210 Watts.

 

[Kocicak]

Stock 210 Watts.

How do you know that, did you run the test? Is this the real power draw or just spec?

 

[Thibsie]

How do you know that, did you run the test? Is this the real power draw or just spec?

LOL. Feel threatened ? xD

 

[Markfw]

How do you know that, did you run the test? Is this the real power draw or just spec?

You can't overclock/underclock or change voltage on EPYC. Why question this ?

 

[Kocicak]

Why threatened? In efficiency alone this server CPU cannot beat my 5700G CPU.

BTW both "cheap" 7950X and 13900K for the consumer platforms can reach 40000 points (13900K can do 36400 at 210W), so 50000 points is not that impressive anymore.

 

[nicalandia]

Why threatened? In efficiency alone this server CPU cannot beat my 5700G CPU.

BTW both "cheap" 7950X and 13900K for the consumer platforms can reach 40000 points (13900K can do 36400 at 210W), so 50000 points is not that impressive anymore.

You are day Dreaming if you believe the 13900K is doing 40K with only 210 Watts

 

[Kocicak]

You are day Dreaming if you believe the 13900K is doing 40K with only 210 Watts

I myself measured 36400 at 210W (I did a lot of power draw imited tests with the 13900K I got early), and that was without undervolting. These CPUs are pretty efficient if you get them out of the insane power draw region.

 

[Markfw]

You are day Dreaming if you believe the 13900K is doing 40K with only 210 Watts

from videocardz:

"he same CPU with power cap unlimited (PL1/PL2 levels at 4095.9W), the single-core performance isn’t better, but multi-core score is now at 40616 points so an improvement of 13.8%. Such an improvement comes at a cost of 100W higher power (345W peak) in the same test. "

Here:

Core i9-13900K breaks 40K points in Cinebench R23 with unlimited power and 5.8 GHz clock - VideoCardz.com

Intel Core i9-13900K compared with and without power limiting settings Leaker OneRaichu shares new benchmark results featuring the flagship Raptor Lake CPU. The 24-core and 32-thread Raptor Lake processor scores 2290 points in single-core Cinebench R23 test and 35693 points in multi-core. It...

videocardz.com

 

[nicalandia]

I myself measured 36400 at 210W (I did a lot of power draw imited tests with the 13900K I got early), and that was without undervolting. These CPUs are pretty efficient if you get them out of the insane power draw region.

50,000 Points for 210 Wat 32 cores with an all core base of 2.6 is really not that bad. It speaks volumes for the Upcoming Zen4 based TR CPUs, I believe they will be breaking 65,000 points stock.

 

[nicalandia]

If you think 50,000 Points its slow for 32 Cores worth of EPYC, then 52,000 points for 56 for Xeon Sapphire Rapids sounds Terrible bad right?

 

[Timmah!]

50 000 points at what power draw? At 150W that would be very impressive. 50 000 points alone mean nothing.

I would say more importantly, what price. I know there are other considerations like the amout of RAM the platform allows (2TB for workstation vs 128GB or 256GB or whatever is the limit on client), but 20 percent performance increase (50000 vs 40000) for like 7x higher price is not a good look in my books.

 

[nicalandia]

I would say more importantly, what price. I know there are other considerations like the amout of RAM the platform allows (2TB for workstation vs 128GB or 256GB or whatever is the limit on client), but 20 percent performance increase (50000 vs 40000) for like 7x higher price is not a good look in my books.

Well, two Xeon Platinum 8450H worth $9400(4,700 a pop) gets you 52,500 points. While a Single 9334 Epyc worth $2,990 nets you 50,000 points.

 

[Timmah!]

Well, two Xeon Platinum 8450H worth $9400(4,700 a pop) gets you 52,500 points. While a Single 9334 Epyc worth $2,990 nets you 50,000 points.

Yeah, for rendering, those are hot garbage at that price. If one has render farm and does stuff for some hollywood studio, then OK, there are surely decent money to be earned to pay for the investment.
But for your 1S home workstation such prices are daylight robbery and make no sense IMO.

Not that i care or want to tell anyone what to buy at what price.

 

[Kocicak]

If you think 50,000 Points its slow for 32 Cores worth of EPYC, then 52,000 points for 56 for Xeon Sapphire Rapids sounds Terrible bad right?

I just checked, 1 Raptor core two threads at 3500 MHz does 1820 points at 7W. 56 of them would make 102 000 points at 400W. If they ran at 2600 MHz, they would make 75 700 points.

There is obviously something wrong with the result.

 

[nicalandia]

I just checked, 1 Raptor core two threads at 3500 MHz does 1820 points at 7W. 56 of them would make 102 000 points at 400W. If they ran at 2600 MHz, they would make 75 700 points.

There is obviously something wrong with the result.

Not at all. You tested Raptor Cove on a Monolithic CPU. This is a OG Golden Cove(Server Class) with 8 tiles(2S), slow Mesh interconnect.

 

[nicalandia]

Yeah, for rendering, those are hot garbage at that price..

For the Xeons? Yes, but keep in mind those are H series with All Tile based Accelerators unlocked. Price wise those will never make sense if you are just rendering. The EPYC 9334 Is actually lower priced than the current 5975WX and just as fast(50K Cinebench),




I suspect that the W9 3495X will be close to Between 65K and 70K in CBR23 Performance.

 

[Timmah!]

For the Xeons? Yes, but keep in mind those are H series with All Tile based Accelerators unlocked. Price wise those will never make sense if you are just rendering. The EPYC 9334 Is actually lower priced than the current 5975WX and just as fast(50K Cinebench),


View attachment 76891

I suspect that the W9 3495X will be close to Between 65K and 70K in CBR23 Performance.

Yeah, for rendering. Obviously i dont comment on server related stuff, that might use those accelerators. Additionally, its not something i have personal experience with, unlike rendering, so it would be stupid to have strong opinions about stuff i do know very little about.
That said, both those EPYC and 5975WX and now SPR Xeons-W are meh as well, even if you are just rendering. If upcoming 32 core Zen4 TR clocks so high, that it provides scores in 70000 range, then we can talk about that one being semi-reasonable.

 

[BorisTheBlade82]

This pretty much confirms that the V-Cache-Dies from Zen3 and Zen4 are NOT the same, as some interpreted from a THG article earlier on.

https://twitter.com/x/status/1627714884648722432

 

[nicalandia]

The 7950X3D Makes its debut on Geekbench5 and Geekbench6

Geekbench5

Geekbench Search - Geekbench

browser.geekbench.com

Geekbench6

Geekbench Search - Geekbench

browser.geekbench.com

 

[JoeRambo]

Geekbench6
Geekbench Search - Geekbench Browser

Not exactly unexpected score. A bit on a low actually for DDR5 6000 + 3D Cache effort.

Score comparison with my 5Ghz locked no/HT 7950x:






In MT interesting is that 32T backed by X3D boosted + 6000 mem cannot beat 16C without HT backed by handtuned 6200 memory. I guess the days of CB23 like scaling in GeekBench are over for good now, lots of fuming inbound.