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!

 

[naad]

I seem to recall quite a number of people calling zen4 "dieshrunk zen3" very satisfied with the performance while core counts stayed the same.
Now if only we could get a <65W SKU that can be used for SFF

 

[moinmoin]

APIKSA Has been delivering the Goods lately so I trust the source.

Sure, I agree. Still 6.5 GHz all core and 7.25 GHz single core are so far beyond anything we have seen before (even previous optimistic rumors) that I prefer staying cautious.

 

[.vodka]

Now if only we could get a <65W SKU that can be used for SFF

Even if there is no low power/non X part available, you can make one yourself that's a good enough replacement even if the silicon hasn't been binned with low power first in mind.

Lower PPT/TDC/EDC to the 65w bracket values (88W/60A/90A)... or enable eco mode. Same thing. Undervolt while you're at it.

Zen4 shines in this power range according to AMD's slides, so performance should be pretty awesome at a fraction of the power.

 

[naad]

Even if there is no low power/non X part available, you can make one yourself that's a good enough replacement even if the silicon hasn't been binned with low power first in mind.

Lower PPT/TDC/EDC to the 65w bracket values (88W/60A/90A)... or enable eco mode. Same thing. Undervolt while you're at it.

Zen4 shines in this power range according to AMD's slides, so performance should be pretty awesome at a fraction of the power.

NUC and SBC like boards don't usually have power control from what I remember, and this wouldn't be a gaming machine so I wouldn't install windows on it, so no ryzen master ( i don't think even laptops have it?)

Linux has some SMU control mechanisms but they're very unofficial, and i don't want to risk blowing something, and BSD has none at all.. so a "T" class SKU would be great, hassle free way to building a 24/7 machine for a fileserver/firewall, or some 35W APUs if AMD ever feels like giving us Rembrandt

 

[igor_kavinski]

The fact that Raphael - and not even Raphael-X - seems to be positioned to dominate Raptor Lake in a large number of benchmarks is not a good look for Intel.

Intel knows that and I'm betting everything I do not own that Intel has something crazy to keep them in the game (like a Z890 platform that accepts a super Raptor Lake on steroids but can be upgraded to Arrow Lake whenever that becomes available).

 

[Kaluan]

Well I gathered the info online, so whatever I found is in the excel sheet. I made an error (clearly) as I assumed that 13400 and 13100 are Raptor Cove based.

Yeah I know most of the info we have now wasn't present back then. Still it's weird seeing people expect 12600K like performance on the 13400. For one, if anyone wants to see 13400 ST perf (and possibly gaming perf as well), seeing how 12500 performs is the easiest way, 'till Q1 2023 when we'll get actual reviews. MT is a different cup of tea tho.

There is no IF... Just look at the L2$ Size.
View attachment 67918

But It's a locked CPU without any boosting technology that has a TDP Cap of 65 Watts.

You should know better than to trust GB in reporting cache sizes... not long ago we had the Genoa GB leak saying it has 16MB per chiplet didn't we?
There's also stuff like this:

Micro-Star International Co., Ltd. MS-7D25 - Geekbench

Benchmark results for a Micro-Star International Co., Ltd. MS-7D25 with an Intel Core i5-12600K processor.

browser.geekbench.com

So unless 12600K is also based on Raptor Lake, I don't see that being conclusive proof of that. Particularly when the leaked 13th gen lineup show 13400 having Alder Lake-like native DDR5 (4800) support as well as literal Intel slides saying non-K i5 and lower only get increased L3 (from the added E-core clusters), only i5K and up get increased L2.

Also there is no such thing as capped TDP, not when it comes to Intel at least, 12900 also has a 65W TDP... but the PL2/28 second boost period is 202W... and most Z690s and some B660 allow for permanent PL2 boosting. So stuff like 12400 are more like 117W parts, not 65W. Expect 13400 to also have a PL2 of 117W (or more).

I seem to recall quite a number of people calling zen4 "dieshrunk zen3" very satisfied with the performance while core counts stayed the same.
Now if only we could get a <65W SKU that can be used for SFF

Well, AMD would be absolute idiots if they didn't also release non-X parts for 35-120W configs, at least for OEM, like they did with 5800 and 5900. Or the 35W 5000GE parts.

IDK, if all else fails, I bet ECO mode has a comeback on AM5/R7000.

 

[eek2121]

I seem to recall quite a number of people calling zen4 "dieshrunk zen3" very satisfied with the performance while core counts stayed the same.
Now if only we could get a <65W SKU that can be used for SFF

You don't really need 65w for SFF. People have been building SFF PCs with 12900ks, 12700ks, 3090s, etc forever. I am actually in the process of building a 7950x system with a 3090 using the FormD T1 v2. Just waiting on the needed parts to launch.

 

[moinmoin]

I'm betting everything I do not own

That's wise of you.

 

[Timmah!]

16 big cores with similar IPC and clocks are always going to win vs a hybrid design like ADL/RL. I expect that intel will come to their senses and step up the number of P cores at some point - if not it will not be pretty.

They will have ample opportunity with Meteor Lake - just place at least 2 of those 6+8 compute tiles onto single chip.

 

[nicalandia]

They will have ample opportunity with Meteor Lake - just place at least 2 of those 6+8 compute tiles onto single chip.

That would be a nightmare for the Thread Director.

 

[igor_kavinski]

That would be a nightmare for the Thread Director.

How about two Thread Directors, each trying to steal threads from the other? That would be cool. Oh, I'm so cold and idle, gimme your threads now!

 

[DisEnchantment]

How about two Thread Directors, each trying to steal threads from the other? That would be cool. Oh, I'm so cold and idle, gimme your threads now!

You will need a senior Director too similar to their corporate hierarchy

 

[nicalandia]

OneRaichu just posted this.... He is tentative on submitting it due to NDA.

Stock Ryzen 9 7950X with DDR6000 RAM

 

[Timmah!]

You will need a senior Director too similar to their corporate hierarchy

Thread Principal

 

[InstrEd]

Thread Principal

As long as it isn't a Nun with a ruler! 😉

 

[InstrEd]

The official reviews can't come soon enough. It does look like AMD has a nice efficient processor to keep the money rolling in.

 

[PJVol]

...this wouldn't be a gaming machine so I wouldn't install windows on it, so no ryzen master ( i don't think even laptops have it?)

Linux has some SMU control mechanisms but they're very unofficial, and i don't want to risk blowing something, and BSD has none at all.. so a "T" class SKU would be great, hassle free way to building a 24/7 machine for a fileserver/firewall, or some 35W APUs if AMD ever feels like giving us Rembrandt

Aside from the mentioned above ECO mode, there's a new driver - amd_hsmp - at least starting from 5.18. It'll allow you to set socket power limit, max boost and socket cclk, turn on/off APB (dynamic DF power managenent), etc. Available HSMP commands are listed in headers /arch/x86/include/uapi/asm/amd_hsmp.h
Besides, there are a couple of good and reliable 3rd party tools providing access to SMU via RSMU (or MP1) mailbox.

PS: the funny thing about RyzenMaster is that it doesn't really follow amd's own recommendation regarding timeout between smu requests, instead it just spam a dozen requests whenever user changes anything, essentially ignoring them.

Same wattage

IMO the energy efficiency comparision of diff. cpu-arch should be made at their respective stock power settings.
I don't think there's a reason to test at the same power cap, let alone any "iso-power"- like nonsense.

 

[dr1337]

NUC and SBC like boards don't usually have power control from what I remember

For BGA based systems yeah this is typically the case but any SFF Zen 4 build is going to be a socketed mini-ITX board with a standard chipset. So power limits and every standard setting will be there. Also any Zen 4 Embedded chips are going to be U or H class, as far as I know AMD has never done a BGA X sku, only HX.

 

[naad]

For BGA based systems yeah this is typically the case but any SFF Zen 4 build is going to be a socketed mini-ITX board with a standard chipset. So power limits and every standard setting will be there. Also any Zen 4 Embedded chips are going to be U or H class, as far as I know AMD has never done a BGA X sku, only HX.

Seems like really going mITX is the way to go, a PCIe slot does wonders for futureproofing if I decide to get a U2 expander or put that good ol' m1015 to use, even tho I had quite some issues with ivybridge gigabyte boards refusing to detect it.. the wonders of server hardware on consumer motherboards I guess?
Just wish at least some ITX AMD boards supported ATX12VO, but even Intel has very few of those so I guess it's too early outside of prebuilds.

Aside from the mentioned above ECO mode, there's a new driver - amd_hsmp - at least starting from 5.18. It'll allow you to set socket power limit, max boost and socket cclk, turn on/off APB (dynamic DF power managenent), etc. Available HSMP commands are listed in headers /arch/x86/include/uapi/asm/amd_hsmp.h
Besides, there are a couple of good and reliable 3rd party tools providing access to SMU via RSMU (or MP1) mailbox.

PS: the funny thing about RyzenMaster is that it doesn't really follow amd's own recommendation regarding timeout between smu requests, instead it just spam a dozen requests whenever user changes anything, essentially ignoring them.

Now that would be nice, kernel supported and enough granuality for me to be satisfied, guess all that's left is seeing how low the new cIOD idles I guess

 

[Just Benching]

IMO the energy efficiency comparision of diff. cpu-arch should be made at their respective stock power settings.
I don't think there's a reason to test at the same power cap, let alone any "iso-power"- like nonsense.

There are 2 reasons that energy efficiency testing should be done in same wattage. First of all, because a fundamental part of electronics is, the increase in wattage isnt linear with the increase in performance. That pretty much means that usually the CPU that has a higher power limit will also have worse efficiency. At that point you are not testing architectural efficiency but just stock settings efficiency. Say a 7950x at 200w is less efficient than a 7950x at 100w.

Second reason, which is pretty similar to the first one, it prevents "cheating".Intel is soon going to realize the 13900T version, which is basically power limited to 35w. That will make it the most efficient cpu on planet Earth, which might not be the case if you tested it in ISO wattage.

Point is, power draw is not an integral part of a CPU. Cores / Cache / frequency etc. are integral parts, power draw isn't. It's the one thing you can change without affecting stability. So there is no reason you shouldn't.

 

[PJVol]

There are 2 reasons that energy efficiency testing should be done in same wattage

1) the fact that power doesn't scale linearly with operating frequency has little to do with the energy efficiency comparison we're talking about here. You probably know that each CPU design is always a compromise between power efficiency and performance, and both manufacturers choose the optimal point on the power / frequency curve, first based on the characteristics of the process node (Vt / Fmax), then the target performance, thermal and electrical design, etc., which is generally a PPT/PL and Fmax limits, that is different for each SKU in a cpu family lineup, let alone different archs and process nodes.
Thus, for the reasons above, an energy efficiency comparison made at the same power limit will obviously be more benefical to one of the parties.

2) I'm not very familiar with modern Intel's power management system, but I'm not sure if this is really an issue.

At that point you are not testing architectural efficiency but just stock settings efficiency. Say a 7950x at 200w is less efficient than a 7950x at 100w.

We can directly compare only 2 models from both lineups, not the architectures as a whole.
So the "same power limit" criterion you suggested only adds another layer of uncertainty to the comparison, for the reasons i stated above.

Cores / Cache / frequency etc. are integral parts, power draw isn't

Can you clarify, how is the frequency is "integral part", and power draw isn't, given that the latter depends on the former, and what this "integral parts" is actually about? )
Core and cache are architecture IPs, and frequency is a certain IP's operation point.

 

[szrpx]

There are 2 reasons that energy efficiency testing should be done in same wattage. First of all, because a fundamental part of electronics is, the increase in wattage isnt linear with the increase in performance. That pretty much means that usually the CPU that has a higher power limit will also have worse efficiency. At that point you are not testing architectural efficiency but just stock settings efficiency. Say a 7950x at 200w is less efficient than a 7950x at 100w.

Second reason, which is pretty similar to the first one, it prevents "cheating".Intel is soon going to realize the 13900T version, which is basically power limited to 35w. That will make it the most efficient cpu on planet Earth, which might not be the case if you tested it in ISO wattage.

Point is, power draw is not an integral part of a CPU. Cores / Cache / frequency etc. are integral parts, power draw isn't. It's the one thing you can change without affecting stability. So there is no reason you shouldn't.

Alright, I'll bite, in what way is the 13900T going to be the most efficient on planet Earth? Getting the most MT performance at that power usage I assume?

 

[nicalandia]

Impressive FP score, but only so-so integer score.

This one is very close to the 2,000 Integer Score at stock with DDR6000 RAM



It's within 5% of the best integer score that the 13900K has posted so far

 

[Kaluan]

After both competitor's new generations are out and reviewed I can already hear casual DIY builders go like...

"wait what!?! 7950X/7900X consumes less power than 13900K/13700K? How is that possible? 170W vs 125W! 😲"

😅

Only reason that didn't happen this gen is because of 105W vs 125W. But you still see people thinking T and non-K 35W/65W Intel parts run at just that wattage in the real world (when it's actually more like 99-202W, based on the SKU)

 

[Just Benching]

1

Alright, I'll bite, in what way is the 13900T going to be the most efficient on planet Earth? Getting the most MT performance at that power usage I assume?

getting the most performance per watt out the box