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!

 

[inf64]

So, I don't find the idea that a 7800X3d could be 30% faster than the 5800X3d in cache sensitive application to be impossible. Its certainly optimistic and likely situational based on application cache sensitivity, but it is definitely possible.

The only problem with your post is that MLID claims Zen 4 with Vcache is 30% faster than Zen 4 without Vcache. So the comparison is not versus Zen 3 Vcache parts. This puts Zen 4 on a whole other performance level.

Food for thought:
IPC gain versus vanilla Zen 3 (conservative as games are memory sensitive and not compute heavy) : 8%
DDR5 gain (conservative in light of how Alder Lake benefits ~ 20%) : 5%
Clock uplift (5.3Ghz boost clock in games versus 4.6Ghz) : ~15%

All combined for vanilla Zen 4 (no Vcache): 1.08 x 1.05 x 1.15= 1.3 or 30% versus regular Zen 3

Now add Vcache effect of around 15% on top and you get these possible numbers for Zen 4 Vcache chip :

1) uplift from Zen 3 : 1.3 x 1.15= 1.5 or 50% faster

2) uplift from Zen 3D : (1.3/1.15) x 1.15= 1.3 or 30% faster

3) uplift from Zen 4 : ~15% faster (provided that game boost clocks are still ~5.3Ghz)

 

[CakeMonster]

I just want to know the 3D lineup and release date so I can plan my purchase accordingly. If it takes too long, or if there won't be a 16c 3D part I might as well buy Z4 on release. I don't really count on AMD telling us as much though...

 

[LightningZ71]

I think that it is possible for zen43d to have a 30% uplift over regular zen4 in certain memory ( but cache effective) bound scenarios. With Zen4 boosting a good 15-20% over what Zen3 was achieving and having a 10% higher IPC, it could easily have 25-30% more demand on its memory subsystem as compared to Zen3. While DDR5 certainly increases bandwidth, it doesn't do much for latency at present. Even with a larger L2, Zen4 chips could easily find themselves more often in a "race to wait" scenario where they blast through their local working set of data and stall on memory reads, having to wait greater and greater percentages of their available work time for a response from main memory. Increasing the L3 size by 200% to 96MB, that could make a dramatic reduction in the percentage of time the core is stalled for memory access, as we see in the 5800x3d. This is amplified in the ZEN4 chips because it is able to churn through data even faster.

 

[Joe NYC]

I don't think anyone would argue that this is not very much in the realm of possibility. The issue is, that's not what he said. He said that a Zen4-3d CPU will perform 30% better than a non-3d Zen4 CPU.

I take the 30% gain to apply to gaming applications, like 15% gain applied to gaming for 5800x3d.

I am kind assuming that Tom misspoke when he said 30% would be overall increase in performance.

 

[Joe NYC]

I just want to know the 3D lineup and release date so I can plan my purchase accordingly. If it takes too long, or if there won't be a 16c 3D part I might as well buy Z4 on release. I don't really count on AMD telling us as much though...

Tom mentioned 7950x3d part, in addition to 7800x3d part.

 

[CakeMonster]

Tom mentioned 7950x3d part, in addition to 7800x3d part.

Not trusting him in any amount whatsoever though. We'll see in a few days I guess.

 

[nicalandia]

I don't think anyone would argue that this is not very much in the realm of possibility. The issue is, that's not what he said. He said that a Zen4-3d CPU will perform 30% better than a non-3d Zen4 CPU.

I can see that happening.

At ISO speed and if they cut down the cycle latency to near 0... Then yeah..! Chips and Cheese made a test and found out that at 0 cycle penalties the performance would be HUGE..

Do IBM’s Giant L3 and V-Cache Represent the Future?

IBM showed off a giant 256 MB L3 during its Telum presentation at Hot Chips 2021, and ignited discussion about whether that represents the future of caches. That’s not the first time we&#8217…

chipsandcheese.com

 

[Zepp]

Why do people even both watching MLID after the BS of his Zen 4 30% IPC "leaks", expect zen 4 -> zen 4 3D to be less of an upgrade compared to zen 3 -> zen 3 3D because DDR5 is going to already help with some of the memory needs that the cache helps with in the 3D version.

sigh* I donno man, I think he's a less agregious clickbaiter than the main tech-tubers (especially LTT and GN). honestly can't stand their channels.
Tom is likable and his broken silicon talks are enjoyable to listen to in the background. Adored was too until he was run off the platform...
...the rumormonger hating has always baffled me though

 

[CakeMonster]

Personally I have much more respect for sloppy benchmarking from people doing actual work, than people who repeatedly make stuff up and then ask for your money.

 

[dnavas]

In the performance comparison between Zen3 and Zen4 AO this is very important information to get across: (Zen4 AO only 5 to 10% faster than Zen3)
View attachment 66529

Without knowing how borked the CPU<-->memory connection was in the Zen4 pre-final silicon test, it's impossible to draw accurate conclusions about a vcache part. I would treat these numbers as "under no circumstances higher than" and assume that MLiD is being used as FUD on the Raptor launch -- particularly as Tom was supposedly supplied with benches.

 

[Joe NYC]

Not trusting him in any amount whatsoever though. We'll see in a few days I guess.

I don't think we will find out details about the V-Cache models during the upcoming Zen 4 announcement - in few days. Probably some time later...

 

[Timmah!]

I don't think we will find out details about the V-Cache models during the upcoming Zen 4 announcement - in few days. Probably some time later...

Lets hope we do. I dont want to wait another half year till March-April for those 3D skus - at that point rumors about next gen (or refresh or whatever) coming out next fall might be already circulating, perhaps increasing the core number or something, and i will be once again inclined to wait... and you never know, at that point maybe even Intel finally puts their crap together and becomes a viable choice. I dont want RPL, but for example i would consider getting 22~24C SPR-X for right price (even higher than Ryzen, but less than TR) over Zen4.

All in all, if they want my money, they better be releasing their 3D stuff ASAP.

 

[Kaluan]

That is because it uses actual workloads. In my experience, it has been pretty accurate at measuring IPC and/or performance and IIRC Dr. Ian Cutress himself stated it was comparable to SPEC.

Real world workloads don’t scale perfectly or linearly.

Geekbench is a solid benchmark that receives way too much hate. Most of that hate comes from the wide variability of results that are published from a wide range of hardware/software configurations ranging from complete garbage to sunshine perfect.

A random result found on their website can be taken with a grain of salt. A result in a carefully controlled environment on the other hand is invaluable for measuring performance.

Yeah I don't know about that, whenever I see a benchmark have results up for something like a 5950X varying from 9K to 17,5K points I can't think of anything but how garbage it is.

Particularly when almost all other benchmarks out there give much more consistent and clear results... and can also be ran in sustained mode, a feature which GB is sorely lacking (given it's mobile SoC focus you'd think throttling/anti-artificial boosting measurements would also be part of it's test suite but no... we're supposed to belive a 30s benchmarks is the be-all end-all of precise performance gauging lol)

And even in ideal scenarios, it's misleading at best. Take 5800X v 12600K results, 5800X typically scores 10xxx points, 12600K scores 11xxxx for DDR4 and 12xxx for DDR5. But taking a quick look at cumulative application performance on review outlets like TPU or ComputerBase, 5800X sits in front of 12600K by a few percent, not 10-20% behind, as GB would make newbies belive.

If anything, I feel like it doesn't get ENOUGH criticism. Still, I hope v6 might pleasantly surprise me one day.

 

[gdansk]

I don't think AMD will announce much about vcache SKUs until nearly ready for availability.
It'd tank sales. Which gamer would buy a 7800X knowing the 7800X3D is coming in 3 months?

 

[Timmah!]

Yeah I don't know about that, whenever I see a benchmark have results up for something like a 5950X varying from 9K to 17,5K points I can't think of anything but how garbage it is.

Particularly when almost all other benchmarks out there give much more consistent and clear results... and can also be ran in sustained mode, a feature which GB is sorely lacking (given it's mobile SoC focus you'd think throttling/anti-artificial boosting measurements would also be part of it's test suite but no... we're supposed to belive a 30s benchmarks is the be-all end-all of precise performance gauging lol)

And even in ideal scenarios, it's misleading at best. Take 5800X v 12600K results, 5800X typically scores 10xxx points, 12600K scores 11xxxx for DDR4 and 12xxx for DDR5. But taking a quick look at cumulative application performance on review outlets like TPU or ComputerBase, 5800X sits in front of 12600K by a few percent, not 10-20% behind, as GB would make newbies belive.

If anything, I feel like it doesn't get ENOUGH criticism. Still, I hope v6 might pleasantly surprise me one day.

Yeah, GB sucks because the variability of results. You bench your rig, go on to look at the comparable results and uh-oh, they are all-over the place, so you have no clue, whether your rig is working as expected or not. Completely misses the point then.
Cinebench on other hand might only be single specific kind of workload, so it should definitely not be taken be all end all either, but at least, from what i could remember, it was always fairly consistent, with the results.

 

[Mopetar]

I'm not sure where 13% came from. Comparing the 5800X3D (4.5 GHz) to the fastest Zen 3 5950X (4.9 GHz) is ~8.8%.

But even if we assumed that we don't lose clock speed with Zen 4 then there should be a larger performance uplift since the 15% from Zen 3D included the 9% lower clock speeds. Simple math suggests that it would make Zen 4D 25% faster than Zen 4 assuming that the extra cache provides the same relative performance gains.

I don't put much trust in MLID, but it's possible that he comes out close to correct on this one anyways.

 

[igor_kavinski]

I don't put much trust in MLID, but it's possible that he comes out close to correct on this one anyways.

If he is correct on that, what about the concerning performance of Zen 4 being on par with 5800X3D? That would mean that in games, Zen 4 is hardly faster than 12900K.

 

[gdansk]

That would mean that in games, Zen 4 is hardly faster than 12900K.

Not surprising. Zen 3 was behind 12900K in games, Zen 4 doesn't improve much IPC but should put it around Alder and Raptor Lake. So it should have similar average performance in games unless Zen 4 clocks higher than 12900K.

 

[Mopetar]

If he is correct on that, what about the concerning performance of Zen 4 being on par with 5800X3D? That would mean that in games, Zen 4 is hardly faster than 12900K.

A person being correct (on accident) in one instance says nothing about their other predictions. I wouldn't add more credibility to those just because he may have stumbled into being close to correct here.

However, there's a problem with your assumption. Zen 4D being 25% (or whatever percent faster it ends up actually being) says nothing about the relationship between Zen 4 and any other processors.

 

[coercitiv]

But even if we assumed that we don't lose clock speed with Zen 4 then there should be a larger performance uplift since the 15% from Zen 3D included the 9% lower clock speeds. Simple math suggests that it would make Zen 4D 25% faster than Zen 4 assuming that the extra cache provides the same relative performance gains.

Exactly the same correction can be applied in reverse to these benchmarks he supposedly saw. Since they were done on early silicon, we can reasonably assume clocks were considerably lower than final target (allowing same clocks on both), we can also assume the DDR5 speed was conservative, maybe even JEDEC specced to their validation target. So even if Zen4D manages to clock closer to equivalent Zen4 silicon, these early tests mean very little without proper context.

 

[Timmah!]

unboxing of Asus X670-E Hero and Gene, if youre interested. In Czech language though. I dont think there is anything new particular being said. But some nice close-up shots.

 

[Mopetar]

That's all well and true, but the 25% I came up with relies on observed data from Zen 3 and 3D and an two assumptions about the effectiveness of infinity cache remaining relatively similar and Zen 4D being able to achieve the same clock speeds as Zen 4.

I didn't base anything off of claims that MLID made, information he was reporting, or anything else in his video. I've merely suggested that it's conceivable that he could wind up being close to correct with a 30% prediction if the assumptions that I've made hold true.

I could easily be wrong about both assumptions. It's certainly possible that the 3D cache chips are still more restricted in how much voltage they can handle, and it's also possible that changes to Zen 4 make the addition of 3D cache less of a performance uplift over the base models.

 

[nader007]

Nice, sounds almost Shakespearean — but AMD, come on, do some proofreading of your slides!

PS. This is also probable evidence that Caps Lock is not good for REAADADEABILITY.

Agree with you.
May be it's time for AMD to learn from INTEL how to make point Slides.
Also a good time for INTEL to learn from AMD how to execute, instead of showing blah ... blah....,
After all INTEL Aurora super computer: Failure.
Sapphire Rapid: Mess.
Ponte vecchio: Failed.
Alchemist cards: Failure.
nano tech Failure.
Only viable tech seem relevant ADL, RPL, which both are power hungry chips.
Go on INTEL,

 

[igor_kavinski]

However, there's a problem with your assumption.

The graph MLID posted possibly has Zen 4 running at lower or similar clocks to 5800X3D. Final silicon may perform significantly better. So I guess nothing to worry about.

 

[Joe NYC]

Lets hope we do. I dont want to wait another half year till March-April for those 3D skus - at that point rumors about next gen (or refresh or whatever) coming out next fall might be already circulating, perhaps increasing the core number or something, and i will be once again inclined to wait... and you never know, at that point maybe even Intel finally puts their crap together and becomes a viable choice. I dont want RPL, but for example i would consider getting 22~24C SPR-X for right price (even higher than Ryzen, but less than TR) over Zen4.

All in all, if they want my money, they better be releasing their 3D stuff ASAP.

For gaming, 7800x3d will be a legendary chip. No regrets...

The most important piece of information coming out of the MLID video is that AMD fixed the voltage problem that held back 5800x3d clock speeds.