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!

 

[Det0x]

Err... did you even bother to look at the last few posts?

It looks like he didn't, because If he did, then he would see the same thing posted by me.

No i was on previous page when posting that

 

[cherullo]

I usually enjoy MLID videos, but he is in a bad streak for sure. The Intel callout is certainly deserved, the excuse that Alchemist needs ReBAR because its memory controller is optimized for large transfers is certainly bollocks. But then Tom makes the crappy argument that Intel couldn't have optimized Alchemist for ReBAR since ReBAR was "launched" way too late in its development cycle, which is totally nonsense too.

His latest Zen4 3D video also had its share of bad logic. You cannot say that the 7800X3D will be 30% faster than the 7800X in gaming based on the presented data because we don't know what are the issues with the A0 stepping. If it has a bad memory controller (or say lower than expected FCLK) then the impact of the added L3 cache will be magnified - just like the performance improvement going from 5800X to 5800X3D is larger when using slower memory, since it hamstrings the 5800X more.

So I still believe that MLID has real sources, but his reasoning skills are faltering lately.

 

[Carfax83]

If these scores are true, then It looks very promising. Why? Because allcore boost shouldn't be only 4.5GHz.
I also have to wonder what was the actual clockspeed during ST test.

edit: 4.5GHz is just the base frequency and not the one this CPU was using during the bench. Just tried my own CPU in It, and It showed the base frequency.

From reports it looks like it was running at 5.4ghz for the single thread test.

 

[deasd]

Allegedly a 7700x es sample

Looks like it's just a 2021 ES

 

[inf64]

If we compare the latest R20 leak with my perf. projection, I nailed it pretty much

Leak : 7700X is 23% faster in ST and 32% faster in MT R20 benchmark

My composite index (bit.ly/3PEY8cM) has 7800X (same 8 core SKU, diff. name) being 24.7% faster than 5800X in ST and 27% faster in MT (global). Note that I assumed 5.5Ghz ST boost and 5.1Ghz MT boost, which seem to basically account for the slight difference in real/projected numbers.

My Zen 3 perf. projection for the whole SKU list was basically spot on. I'm now pretty confident this one will be within a margin of error.

 

[Joe NYC]

Dr. Lisa "sandbagging" Su.

Dr. Lisa ,"unleash the" tSunami

 

[TESKATLIPOKA]

Dr. Lisa ,"unleash the" tSunami

Nah.
Release the Kraken.

 

[Carfax83]

So clock for clock it seems to be slightly behind Alder Lake in terms of single threaded performance....at least in this benchmark. By slightly I mean 5 or 6%. That's pretty good, and better than I expected to be honest.

Other applications may show a higher or lower performance difference.

 

[Asterox]

Based on the latest leak:

View attachment 66545

Maybe, but Cinebench is not ultimate CPU benchmark.

https://twitter.com/x/status/1562480259118686208

Cinebench is just a quick narrowly/rendering focused test.

Page 33 - Zen 2 APUs/"Renoir" discussion thread

Page 33 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[gdansk]

So clock for clock it seems to be slightly behind Alder Lake in terms of single threaded performance....at least in this benchmark. By slightly I mean 5 or 6%. That's pretty good, and better than I expected to be honest.

Other applications may show a higher or lower performance difference.

Do you know what clock rate the 7700X is running at? It says 4.5GHz but Cinebench says the base clock rates frequently.
What does 12900K get at the same clock rate?
Basically, please expand on your post with some numbers.

 

[inf64]

Maybe, but Cinebench is not ultimate CPU benchmark.

https://twitter.com/x/status/1562480259118686208

Cinebench is just a quick narrowly/rendering focused test.

Page 33 - Zen 2 APUs/"Renoir" discussion thread

Page 33 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

That's true, but it's interesting that IPC increase in R20 is pretty much spot on with the average figure AMD provided. I expect that in memory sensitive workloads, Zen 4 will have higher IPC than Golden Cove by a few percent (~5%) so the whole average for ST and MT will even out.

 

[nicalandia]

Do you know what clock rate the 7700X is running at? It says 4.5GHz but Cinebench says the base clock rates frequently.
What does 12900K get at the same clock rate?
Basically, please expand on your post with some numbers.

Zen4 is Not 6% behind Alder Lake clock for clock...

 

[jpiniero]

Do you know what clock rate the 7700X is running at? It says 4.5GHz but Cinebench says the base clock rates frequently.
What does 12900K get at the same clock rate?
Basically, please expand on your post with some numbers.

From AT's review the 12900K at stock settings gets 759/10303.

 

[inf64]

Zen4 is Not 6% behind Alder Lake clock for clock...

Well in the "old" R20 they are pretty close. In ST the Golden cove is around ~4.6% faster assuming 7700X ran at 5.4Ghz (which is the expected turbo for retail part), while in MT the difference is just ~2%. In memory sensitive apps Zen 4 should easily eclipse the Golden/Raptor Cove.

 

[deasd]

Any one notice the MP ratio is so high? 9.97x...... what I can found through web is 9.7x-9.8x for 5800x.

Is Zen4's SMT yield is a bit higher than Zen3's, or this sample of Zen4's clock is weird?

 

[Det0x]

Do you know what clock rate the 7700X is running at? It says 4.5GHz but Cinebench says the base clock rates frequently.
What does 12900K get at the same clock rate?
Basically, please expand on your post with some numbers.

8core GC @ 5ghz static:
ST= 760
MT = 8000

 

[inf64]

Any one notice the MP ratio is so high? 9.97x...... what I can found through web is 9.7x-9.8x for 5800x.

Is Zen4's SMT yield is a bit higher than Zen3's, or this sample of Zen4's clock is weird?

It seems that MT IPC is a bit higher for Zen 4 vs Zen 3. Looks like ~9-10% for MT IPC versus ~7-8% for ST IPC (in workloads like Geekbench or Cinebench, at least).

Here is a good post from HXL on twitter. It's showing a difference in instruction mix for a typical modern PC game vs Cinebench R20 (game basically has 97% non-FP ops):

https://twitter.com/x/status/1562485396461670401

In summary, the R20/Geekbench IPC jump will be a lot lower than performance in integer heavy workloads (something that angstronomics hinted at before):

"As for the claimed Single-Thread Uplift of ‘greater than 15% expected‘, Angstronomics can confirm this is a conservative value, done at below final frequencies and using Maxon’s Cinebench R23 Single Thread Benchmark. We can independently confirm that the Performance Per Clock (PPC) targets for the Zen4 core are targeted at +7% Single-Thread PPC, +10% Multi-Thread PPC over their Zen3 core, with significantly higher PPC for memory sensitive workloads thanks to DDR5 while core execution bound workloads like Cinema4D have a lower PPC improvement. "

 

[nicalandia]

Any one notice the MP ratio is so high? 9.97x...... what I can found through web is 9.7x-9.8x for 5800x.

Is Zen4's SMT yield is a bit higher than Zen3's, or this sample of Zen4's clock is weird?

The MP Ratio is High for the same reason the MT on Zen4 is going to be >35%... Due to Higher All Core Boost this time around. In the past Zen3 Was limited by TDP(at stock).. When you enabled PBO the MP ration goes up...

For example

 

[TESKATLIPOKA]

8core GC @ 5ghz static:
ST= 760
MT = 8000
View attachment 66549

And package power for that is how much? ~250W?

 

[Markfw]

The other thing I want to see is what memory speed they are using in all these tests, both the 13900k and the 7950x. I would love to see matched memory at 6000-6600 speed testing on those 2.

Edit: and power usage on a long term (>5 minutes) test.

Edit2: And on no more than at most a 360 AIO for cooling.

 

[Det0x]

And package power for that is how much?

Unknown, i'm running a 5950x inching to upgrade to a 7950x
CPU in screenshot is from a user over at overclock.net forum running a binned cpu and memory optimized @ 6400MT/s together with uncore at 4700mhz.
(not that memory performance dont matter much in Cinebench anyway)

 

[nicalandia]

The other thing I want to see is what memory speed they are using in all these tests, both the 13900k and the 7950x. I would love to see matched memory at 6000-6600 speed testing on those 2.

Edit: and power usage on a long term (>5 minutes) test.

Edit2: And on no more than at most a 360 AIO for cooling.

Overall the 7950X will be the better CPU for Workstation workloads, first due to being 16 Big Cores vs 8 +16, the Zen4 will have AVX512 enabled and surely the MB will be a longed lived one. Raptor Lake is the end of the line for that Z690 Motherboard. You will not see Meteor Lake on Desktop and Arrow Lake is a 2024 product(Optimistically)

 

[Markfw]

Overall the 7950X will be the better CPU for Workstation workloads, first due to being 16 Big Cores vs 8 +16, the Zen4 will have AVX512 enabled and surely the MB will be a longed lived one. Raptor Lake is the end of the line for that Z690 Motherboard. You will not see Meteor Lake on Desktop and Arrow Lake is a 2024 product(Optimistically)

But the trolls will only talk about how it beats the 7950x on single core (if it does) and lightly threaded or short duration tasks. I am betting that the 7950x will win in everything but single core, and close on that one. And I bet the power usage in all scenarios will trash the 13900k.

 

[tamz_msc]

But the trolls will only talk about how it beats the 7950x on single core (if it does) and lightly threaded or short duration tasks. I am betting that the 7950x will win in everything but single core, and close on that one. And I bet the power usage in all scenarios will trash the 13900k.

So the people for whom the short duration performance is more important are trolls?

 

[nicalandia]

So the people for whom the short duration performance is more important are trolls?

A single Run of MT Cinebench is really not applicable to people that actually use their computer for work. Hence a 10 minute CB Run the Intel CPUs lose more performance due to thermals and power budget when compared to AMD

Alder Lake lose more than 10% MT performance on a 10 minute loop as compared to less than 1% of the AMD top of the line. This will translate linearly to Raptor Lake.

So say a stock 16C/32T 7950X beats/matches(by a 1-2% at stock) a 13900K/S, You know that for work the 7950X will actually perform about 10% better due to thermals/Power budget and that Intel is already pushing those CPUs to the limit.