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!

 

[igor_kavinski]

It's just those two benchmarks that look to be off.

I guess the next logical thing for them would be to investigate by turning off the E-cores. Will they do that? My twit account is blocked coz they think I'm some bot. Maybe you could tweet to HWU and ask them to investigate?

 

[Carfax83]

I guess the next logical thing for them would be to investigate by turning off the E-cores. Will they do that? My twit account is blocked coz they think I'm some bot. Maybe you could tweet to HWU and ask them to investigate?

Well I would, if I had a twitter account I think I'll reach out to them in their YouTube comment section and hope it gains enough traction to warrant a response.

 

[Kaluan]

I'm not saying the 7600x can't be faster than the 13600K in gaming. Don't get me wrong. What I'm saying is that HWU's benchmarks for HZD and BF5 are clearly anomalous and need further looking into.

It's just those two benchmarks that look to be off.

AMD uArch performs really good in title X or Y = ANOMALOUS! Must disregard result/investigate/tweak Intel uArch so it's "fair"!

Got it. lol

Where were you when Hardware Unboxed started benching The Riftbreaker (a game barely anyone knows or paid attention to) during Alder Lake's launch? A game OFFICIALLY optimized for Intel CPUs, by the developer's direct admission.

That's not anomalous, right? 🤣

 

[Carfax83]

AMD uArch performs really good in title X or Y = ANOMALOUS! Must disregard result/investigate/tweak Intel uArch so it's "fair"!

Got it. lol

Where were you when Hardware Unboxed started benching The Riftbreaker (a game barely anyone knows or paid attention to) during Alder Lake's launch? A game OFFICIALLY optimized for Intel CPUs, by the developer's direct admission.

That's not anomalous, right? 🤣

I don't even know why you brought up the Riftbreaker as they are in no way comparable.

The Riftbreaker isn't displaying the same kind of performance gap that BF5 and HZD are displaying in the latest HWU review. This is the Riftbreaker benchmark from the 13900K review. Do you see anything weird?

The game seems well optimized for both Intel and AMD and is actually being bottlenecked by the RTX 4090 at 1080p as it's both CPU and GPU intensive.

Compared to HZD:

 

[Doug S]

And as I said before, you double the CPU wattage, and it costs you 6x. 2x cpu power, 2x cpu cooling, 2x cpu generator backup. And now with power at a premium, that 4x for power is like 16x. Things WILL change in the data centers, or people ARE going to get fired.

I've consulted on more than one datacenter design & build out. Those numbers are a complete fantasy, you have no clue what you are talking about.

Even if CPUs were responsible for 100% of a datacenter's power consumption, and storage, memory, networking, conversion overhead, etc. etc. used zero energy those numbers would be ridiculous.

 

[igor_kavinski]

I've consulted on more than one datacenter design & build out.

But you do agree that Intel server CPUs working full tilt on mission critical workloads in high density server racks would be much harder to keep cool than AMD Epyc's, right?

 

[Kaluan]

I don't even know why you brought up the Riftbreaker as they are in no way comparable.

The Riftbreaker isn't displaying the same kind of performance gap that BF5 and HZD are displaying in the latest HWU review. This is the Riftbreaker benchmark from the 13900K review. Do you see anything weird?

The game seems well optimized for both Intel and AMD and is actually being bottlenecked by the RTX 4090 at 1080p as it's both CPU and GPU intensive.

Compared to HZD:

Were did I say The Riftbreaker was about today's CPU lineup? I specifically said during ADL launch window. Reading comprehension?

You know... this:


They (EXOR Studios) had a blog on Intel's website saying how they're optimizing for Intel (read: Golden Cove). The fact that Zen4 caught up in that game (kudos to AMD) has nothing to do with what I initially said.

Don't think you were white knight-ing for AMD back then, were you? You just took The Riftbreaker results as-is: A great result for Intel (optimizations aside).

And for crying out loud, Horizon ZD performed great on Ryzen 5000 as well, it was one of the arguably few games were Zen3 bested Golden Cove.

We even knew from the early Chinese QS/ES RTL v RPL gaming review that Zen4 will probably do very well vs Raptor Lake in that game. HU's results aren't exactly shocking. Especially w/ a 4090, were leads may grow vs w/ 3090Ti or 6900XT (notice how old results have 5950X = 12900K+fast DDR5, but in the ones w/ 4090, 5950X puls a bit ahead)

This is just ridiculous nitpicking TBH.

 

[fkoehler]

Last 20 years has been all with F500 big hitters at the top of the Med/Bio-tech.
TCO is always either tied with #1 or is #2 on any list as long as its fit for purpose.
No one thinks people are going to get fired for not replacing their infrastructure. However everyone has known for a good 6 months that power is going to go up significantly, with significant impact on TCO.
If you're not updating your standard model, or replacing EOL systems with an equivilent/cheaper AMD that also has a lower watts/btu for lower TCO, I think it will at least give you a painful ding on that performance review.
I used to do big CSR's, and a few times we had Cisco come out with new models that had a particular feature that was interesting enough that we re-did the CSR. Budgets are set out at the beginning of the year, however its not like you spend your entire budget on the first of the fiscal year.
Might have been a bit hyperbolic in my previous quote, however AMD being basically at least equivilent performance-wise, and power-wise sure seems like a reason for AMD's Server business to increase.

Not gonna happen, that's not how things are done in big corporations in my (rather substantial) experience. Capital spending for datacenter grade servers is budgeted yearly and typically part of a longer term multi year plan. They know well in advance which servers they will be replacing as they reach the end of their lifecycle and/or are fully depreciated, and know how much they have budgeted to spend on new ones. They can't just decide to replace servers early because of temporary changes in energy prices, and qualification of new platforms typically takes months so they aren't even close to considering Zen 4 when it has been out for what a week now (though many will have had test systems for a few months already)

 

[Carfax83]

Were did I say The Riftbreaker was about today's CPU lineup? I specifically said during ADL launch window. Reading comprehension?

You know... this

So you're mad because Alder Lake was significantly faster than Zen 3 in Riftbreaker? Newsflash, Alder Lake was newer than Zen 3 (which had already been out for a year) and possessed the advantages of DDR5, higher clock speeds and higher IPC. You would have to be firmly in zealot territory to see anything wrong with that.

It's clear that Riftbreaker loves cache, clock speed and bandwidth and Alder Lake provided all three in healthy amounts which is why it performed so well on 12th gen. The 5800x 3D clawed back a lot of performance due to its humongous cache and is a whopping 35% faster than the 5800x in Riftbreaker for instance.

They (EXOR Studios) had a blog on Intel's website saying how they're optimizing for Intel (read: Golden Cove). The fact that Zen4 caught up in that game (kudos to AMD) has nothing to do with what I initially said.

Optimizing in this case likely meant making sure the E cores weren't being used by the game, or using them for background tasks. Hitman 3 did the same thing. It had nothing to do with optimizing specifically for Intel architectures. As I said before, PC developers don't target particular CPUs because the PC is an open platform with tens of thousands of permutations, or millions if you count non gaming PCs.

And this is shown by how much better Zen 4 did compared to Zen 3 without any specific optimization on the developer's part for Zen 4, because Zen 4 has higher IPC, clock speeds, more cache and more bandwidth; all the things that the game loves.

Don't think you were white knight-ing for AMD back then, were you? You just took The Riftbreaker results as-is: A great result for Intel (optimizations aside).

I've been PC gaming for over 20 years now and I know a decent amount about how 3D engines work and how they leverage hardware. When I first saw Riftbreaker benchmarks, I did some research to see why it performed so well on Alder Lake in comparison to not just Zen 3, but other CPUs.

Seeing the game in action told me most of what I needed to know. The game is actually very technically sophisticated and supports RT shadows and ambient occlusion, as well as a sophisticated physics engine for destruction, particle effects and vegetation deformation. The game also features thousands of entities on screen at once.

All these things together makes the game very CPU demanding, so CPUs that can process and dispatch more instructions have the advantage. Alder Lake possessed these qualities in greater measure comparative to Zen 3, as it's a much wider architecture with more instruction throughput and bandwidth. These qualities were sustained for Raptor Lake, as Raptor Lake is also faster than Zen 4 in Riftbreaker for the same reasons.

In fact, the 13900K is still being GPU bottlenecked by an RTX 4090 at 1080p in this game.

And for crying out loud, Horizon ZD performed great on Ryzen 5000 as well, it was one of the arguably few games were Zen3 bested Golden Cove.

Yes, and I always found that strange. I expect things like that to happen with graphics cards but not with CPUs because developers target x86-64 for CPUs and the CPUs using their internal OoO logic and microarchitecture determine how fast those instructions are processed. On paper, Alder Lake has a distinct advantage over Zen 3 in all the performance categories ie clock speed, IPC, instruction throughput, bandwidth etcetera.

It should theoretically blow the doors off of Zen 3 in HZD. It's also weird that in their 12900KS review, Alder Lake is in front of Zen 3, but in the 13900K review, it's behind Zen 3.

Try explaining that one:

This is just ridiculous nitpicking TBH.

What you call nitpicking I call asking questions. Can you explain the above graph, where the 12900KS is in front of the 5950x, but in the 13900K review (a faster CPU than the 12900KS), it's behind the 5950x?

Perhaps there are sections in the game that perform better on Intel or on AMD.

 

[alexruiz]

Hold your horses. That BFV benchmark he did obviously has something wrong with it as it doesn't make sense. There is no way on God's green Earth that the 7600x is 36% faster than the 13600K in BF5 and then 6% slower in BF 2042.

Steve should have thrown out that result as it's not possible. Those kinds of extreme outliers are obviously incorrect. I would say the same for Horizon Zero Dawn and some of the others. It might be a problem with the efficiency cores getting in the way, but there is no way the 7600x leads by such a huge margin in those games.

And just looking at the game selection, the 13600K is more dominant in CPU demanding games. Call of Duty MW2 MP performs better on the 13600K compared to the single player game as the MP is more CPU demanding. Raptor Lake has a stronger core than Zen 4 in terms of raw CPU power.

I usually like HWU, but no one else is coming out with these weird results and publishing them.

I don't think you get it
As I wrote in several other posts "A lot of the so called reviewers don't have a clue on how to configure a Zen 4 system. The numbers for RPL have little variance among the reviewers, the numbers for Zen 4 have huge discrepancies among them all"
The numbers for RPL will not change that much, the numbers for Zen4 will keep getting better and better as more reviewers get the hang on how to tune the systems.

Techspot is probably the one that tuned the Zen4 system the best, and it is not even about exotic hardware or extenuating memory timings. Some very basic things have major impact:
- Latest UEFI. Always important, specially critical on as new platform. There is quite a jump in sustained performance from AGESA ComboAM5 1.0.0.1 to 1.0.0.3A with enhanced curve optimizer.
- Chipset drivers. Extremely overlooked, and I do not think any of the youtubers bother to install them, but these are also critical. The chipset drivers install Ryzen power plans which will keep the windows scheduler loading the faster cores as preferred cores. Don't install them and have windows bounce between cores and CCDs and tank performance.
- EXPO/ XMP: Even if the timings are not optimal, going from JEDEC DDR5-4800 CL48 to XMP DDR5-6000 CL36 will help a lot.
- Fresh windows install: Are they deploying all the systems from a single master image, or from an intel master and an AMD master? I can tell you that most youtubers will use a single master image, and more than likely that image was created on an intel system.

So, answer me these questions.
Your preferred outlets, those that are showing the results YOU like, how are they configuring the Zen4 systems?
What motherboard? What UEFI version?
Did they install AMD chipset drivers? If so, as they should, what version?
What windows version? What updates version? Fresh for AMD?
I can bet money that your preferred reviewers didn't bother to flash the latest UEFI, didn't bother to install chipset drivers, and no way they had a dedicated AMD windows master image.

On your last pointy, about RPL being a faster core, I don't know.
The 2MB L2 per core helped it a lot, it is also boosted very aggressively, but on pure raw core performance? I don't know

 

[Carfax83]

I don't think you get it
As I wrote in several other posts "A lot of the so called reviewers don't have a clue on how to configure a Zen 4 system. The numbers for RPL have little variance among the reviewers, the numbers for Zen 4 have huge discrepancies among them all"
The numbers for RPL will not change that much, the numbers for Zen4 will keep getting better and better as more reviewers get the hang on how to tune the systems.

Dude, it's not rocket science. Reviewers do this stuff for a living, and AMD told them exactly how to configure it to get the best performance.

Techspot is probably the one that tuned the Zen4 system the best, and it is not even about exotic hardware or extenuating memory timings. Some very basic things have major impact:

Yes, how very convenient. And I'm not saying that the issue is necessarily Tech Spot's testing methodology. The game itself may be bugged for all we know. How else do you explain how a 5950x beat a 13900K with DDR5?

- Latest UEFI. Always important, specially critical on as new platform. There is quite a jump in sustained performance from AGESA ComboAM5 1.0.0.1 to 1.0.0.3A with enhanced curve optimizer.
- Chipset drivers. Extremely overlooked, and I do not think any of the youtubers bother to install them, but these are also critical. The chipset drivers install Ryzen power plans which will keep the windows scheduler loading the faster cores as preferred cores. Don't install them and have windows bounce between cores and CCDs and tank performance.
- EXPO/ XMP: Even if the timings are not optimal, going from JEDEC DDR5-4800 CL48 to XMP DDR5-6000 CL36 will help a lot.
- Fresh windows install: Are they deploying all the systems from a single master image, or from an intel master and an AMD master? I can tell you that most youtubers will use a single master image, and more than likely that image was created on an intel system.

This is all just basic stuff. I think you are severely underestimating the competence of most YouTube reviewers.

So, answer me these questions.
Your preferred outlets, those that are showing the results YOU like, how are they configuring the Zen4 systems?
What motherboard? What UEFI version?
Did they install AMD chipset drivers? If so, as they should, what version?
What windows version? What updates version? Fresh for AMD?
I can bet money that your preferred reviewers didn't bother to flash the latest UEFI, didn't bother to install chipset drivers, and no way they had a dedicated AMD windows master image.

HWU is one of my preferred go to channels for reviews, along with Computerbase.de and a few others. I always thought they did a bang up job and I still do. But I think they do themself a disservice when they publish benchmarks that are nonsensical without additional information or context, because it makes people wonder if they made a mistake or worse, that they are biased because they choose to upload those graphs knowing they are misrepresentative.

On your last pointy, about RPL being a faster core, I don't know.
The 2MB L2 per core helped it a lot, it is also boosted very aggressively, but on pure raw core performance? I don't know

Zen 4 is a 4 issue CPU, while Raptor Lake is a 6 issue CPU. Raptor Lake also has a bigger OoO window than Zen 4, 512 vs 320. Raptor Lake can also do 3x 256 bit loads per cycle vs 2 with Zen 4. These are just a few of the differences. I'm not saying that Zen 3 or Zen 4 is weak by any means, but Golden and Raptor Cove are just significantly wider with higher throughput.

Case in point, Zen 4 never trounced Alder Lake when it launched. Alder Lake got plenty of wins in despite having a lower clock speed than Zen 4.

 

[inf64]

Zen 4 is a 4 issue CPU, while Raptor Lake is a 6 issue CPU. Raptor Lake also has a bigger OoO window than Zen 4, 512 vs 320. Raptor Lake can also do 3x 256 bit loads per cycle vs 2 with Zen 4. These are just a few of the differences. I'm not saying that Zen 3 or Zen 4 is weak by any means, but Golden and Raptor Cove are just significantly wider with higher throughput.

All those beefed up resources for Raptor Cove, dare I say wasted, to get effectively the same IPC as Zen 4 (as per SPEC_1T tests). This goes to show that bruteforcing your design to cover inefficiencies elsewhere is not the best way to go . I'm impressed how AMD managed to get to the same IPC with their clever engineering, instead of spewing resources into ROB (60% difference!).

Raptor Lake is a good chip considering how intel was limited to the same process node, and its gaming performance is where it is due to it being monolithic (slightly faster than Zen4, on average). This will all be mitigated once Zen4 with Vcache launches, then intel will have to again one up AMD with Meteor Lake.

 

[DrMrLordX]

intel will have to again one up AMD with Meteor Lake.

Probably not happening.

 

[igor_kavinski]

The game itself may be bugged for all we know. How else do you explain how a 5950x beat a 13900K with DDR5?

The game versions could be different between the two reviews. Windows versions will likely be different. GPU is definitely much more powerful in the 13900K review. Last, but not the least, graphics drivers were tuned for the 4090 (probably to parallelize them more to keep the 4090 fed with data) and we all know that AMD CPUs are good at multicore workloads, especially when multitasking (Ganesh's 5800H review showed Zen 3's foreground thread suffered lower performance loss than Alder Lake when compute intensive background tasks were running).

While I think this situation warrants further investigation, I don't think this is the result of some "bug". This is how things will be going forward, unless HZD developers put in significant effort to re-tune their engine for Raptor Lake with a separate codepath or optimized executable.

 

[igor_kavinski]

Probably not happening.

If the rumors of the delay are true, Intel will be lucky to have something match Zen 5, let alone beat it.

 

[DrMrLordX]

If the rumors of the delay are true, Intel will be lucky to have something match Zen 5, let alone beat it.

All that aside, Meteor Lake will likely come with clock regression and be targeted at mobile. The rumoured Raptor Lake Refresh will be the competition for Raphael-X.

 

[In2Photos]

https://twitter.com/x/status/1591802741265166336

Raptor cove showing its brutal horsepower 😏

Notice who liked this post but didn't say anything about it being anomalous.

 

[Carfax83]

All those beefed up resources for Raptor Cove, dare I say wasted, to get effectively the same IPC as Zen 4 (as per SPEC_1T tests). This goes to show that bruteforcing your design to cover inefficiencies elsewhere is not the best way to go . I'm impressed how AMD managed to get to the same IPC with their clever engineering, instead of spewing resources into ROB (60% difference!).

I've said it before many times, but I'm not a fan of Spec. To me Spec doesn't reflect real world performance very well. Zen 4 never decisively beat Alder Lake despite having a clock speed advantage of 400 to 500mhz which to me indicates that Alder Lake still had an IPC advantage.

Raptor Lake is a good chip considering how intel was limited to the same process node, and its gaming performance is where it is due to it being monolithic (slightly faster than Zen4, on average). This will all be mitigated once Zen4 with Vcache launches, then intel will have to again one up AMD with Meteor Lake.

To me Raptor Lake is almost a perfect desktop CPU. It's good that you are giving Intel some credit at least, particularly about being on the same process node. One thing where Intel should get their due credit as well is that they managed to nearly double the read bandwidth of the L3 cache in Raptor Lake.

 

[Carfax83]

The game versions could be different between the two reviews. Windows versions will likely be different. GPU is definitely much more powerful in the 13900K review. Last, but not the least, graphics drivers were tuned for the 4090 (probably to parallelize them more to keep the 4090 fed with data) and we all know that AMD CPUs are good at multicore workloads, especially when multitasking (Ganesh's 5800H review showed Zen 3's foreground thread suffered lower performance loss than Alder Lake when compute intensive background tasks were running).

The game version, graphics drivers are all legitimate concerns, but as far as parallelization is concerned, the HZD engine cannot scale to 16 threads let alone 32. Very few 3D engines can scale that high. The only one off the top of my head is the IdTech 7 engine that Doom Eternal uses and it doesn't have a lead rendering thread. The vast majority of 3D engines can only scale to 6 to 8 threads, even with DX12 or Vulkan. An engine has to be specifically built from the ground up to use that many threads. .

While I think this situation warrants further investigation, I don't think this is the result of some "bug". This is how things will be going forward, unless HZD developers put in significant effort to re-tune their engine for Raptor Lake with a separate codepath or optimized executable.

It could also be just a specific area in the game that is favorable to Zen CPUs. Who knows? But to me the difference is too large to indicate such a problem.

 

[Carfax83]

Notice who liked this post but didn't say anything about it being anomalous.

I did that intentionally knowing that one of you guys would pick up on it. Thanks for falling into my trap!

All jokes aside, that part of the game really pushes CPUs. Several game reviewers mentioned that they took big hits to their framerate in the town areas, or when the rat swarm was on screen. Raptor Lake has more CPU grunt than Zen 4 though, so I'm not surprised it manages to beat it in that area. It's far less surprising than Zen 3 beating Raptor Lake in HZD in HWU's review I'll tell you that!

I'm tempted to buy the game myself just to test it as it seems to truly be a next gen tittle in many respects.

 

[deasd]

R9-7845HX

GPU Benchmark - Ashes of the Singularity

Video Card GPU Benchmarks, High quality testing of your video card and gaming system.

www.ashesofthesingularity.com

from the new naming scheme we could easily know this is Zen4 mobile 12 cores

 

[inf64]

I've said it before many times, but I'm not a fan of Spec. To me Spec doesn't reflect real world performance very well. Zen 4 never decisively beat Alder Lake despite having a clock speed advantage of 400 to 500mhz which to me indicates that Alder Lake still had an IPC advantage.



To me Raptor Lake is almost a perfect desktop CPU. It's good that you are giving Intel some credit at least, particularly about being on the same process node. One thing where Intel should get their due credit as well is that they managed to nearly double the read bandwidth of the L3 cache in Raptor Lake.

It's not just spec, it's Geekbench as well - Dr Cutress mentioned that Geekbench was a good IPC measuring tool (for general purpose IPC). Zen 4 and Raptor Cove are basically neck and neck in GB5(even with cryptography excluded; intel part runs at 100Mhz higher clock in ST test): https://www.anandtech.com/show/17601/intel-core-i9-13900k-and-i5-13600k-review/12

 

[In2Photos]

I did that intentionally knowing that one of you guys would pick up on it. Thanks for falling into my trap!

All jokes aside, that part of the game really pushes CPUs. Several game reviewers mentioned that they took big hits to their framerate in the town areas, or when the rat swarm was on screen. Raptor Lake has more CPU grunt than Zen 4 though, so I'm not surprised it manages to beat it in that area. It's far less surprising than Zen 3 beating Raptor Lake in HZD in HWU's review I'll tell you that!

I'm tempted to buy the game myself just to test it as it seems to truly be a next gen tittle in many respects.

Right, so when the game favors Intel it's because the game really taxes the CPU, but when it favors AMD it has to be a glitch.

 

[eek2121]

It's not just spec, it's Geekbench as well - Dr Cutress mentioned that Geekbench was a good IPC measuring tool (for general purpose IPC). Zen 4 and Raptor Cove are basically neck and neck in GB5(even with cryptography excluded; intel part runs at 100Mhz higher clock in ST test): https://www.anandtech.com/show/17601/intel-core-i9-13900k-and-i5-13600k-review/12

Crypto is capped at a low percentage of the score btw. Furthermore, it is a valid workload just like the others. It shouldn’t be excluded IMO.

People just gawk at a seemingly large number and think it is artificially inflating the score, or that the workload ONLY represents cryptography. Neither is true.

 

[trexfromouterspace]

To be honest, the only workload that matters is Elden Ring