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!

 

[MadRat]

What stops AMD from simply remapping vcache at the system level? Instead of throwing 32MB of L3 evenly across 16 cores, what stops them from adjusting it on the fly? If they wanted to throw 32MB at core0, and withhold the rest from core1--> core15, that could really mess with benchmarks. Likewise, a core having little benefit from L3 shouldn't hold any hostage from the others.

 

[poke01]

Ugh, that guy spends more time posting bait tweets that are either pro-Intel or pro-Nvidia rather than tweets that promote his own software.

Also, I don't understand people's self-proclaimed responsibility for having to "keep AMD hype in check", especially when I don't really see the market as being heavily AMD favored to begin with, and not when AMD is still the underdog in terms of company size. I refuse to believe that a handful of Youtube leakers constitute the bastion of "over hyping" that he's trying to "correct".

Most of the hype that leakers say about AMD that is insane to begin with is about their upcoming GPUs.

"RDNA 3 will have better efficiency than Ada and performance too"

"91 TFLOP for Navi 31"

I hate leak culture for AMD. Nvidia leak hype is just to get negative clicks. "600 watt GPU" turns out the RTX 4090 is very efficient for the power it offers.

Also AMD is not an underdog anymore. They beat Intel in server CPUs and laptop/Ultrabooks when it comes to efficiency and actually release most of their products on time.

Their Market cap is now around 100 Billion and in 2015 it was 1 Billion.

AMD is not an underdog anymore it is a premium CPU brand and a high end GPU brand.

Nvidia is an high end/ultra high end GPU brand and Intel is the delay brand.

 

[Kocicak]

i did not fancy little cores shenanigans

Having little cores on a CPU is like having a bus full of handy gnomes with spades, they just rush out and get to work, you have to love them!

 

[Timmah!]

Having little cores on a CPU is like having a bus full of handy gnomes with spades, they just rush out and get to work, you have to love them!

sorry, but meh. Would take 24c 2495x, if it was priced right, over “24c” 13900k, even if the latter is faster in single thread.

 

[In2Photos]

Having little cores on a CPU is like having a bus full of handy gnomes with spades, they just rush out and get to work, you have to love them!

Would you feel the same way if AMD was using little/big and Intel had big cores only?

 

[Timmah!]

Most of the hype that leakers say about AMD that is insane to begin with is about their upcoming GPUs.

"RDNA 3 will have better efficiency than Ada and performance too"

"91 TFLOP for Navi 31"

I hate leak culture for AMD. Nvidia leak hype is just to get negative clicks. "600 watt GPU" turns out the RTX 4090 is very efficient for the power it offers.

Also AMD is not an underdog anymore. They beat Intel in server CPUs and laptop/Ultrabooks when it comes to efficiency and actually release most of their products on time.

Their Market cap is now around 100 Billion and in 2015 it was 1 Billion.

AMD is not an underdog anymore it is a premium CPU brand and a high end GPU brand.

Nvidia is an high end/ultra high end GPU brand and Intel is the delay brand.

Nvidias aggressive and deceiving marketing is indeed deplorable, but i can attest the pro AMD fanboys among the leakers and so called “journalists” are equally annoying and terrible. There is very pro AMD local web, the main writer and admin likes to tell about himself how “objective” he is and no fan of any brand, when the exact opposite could not be more obvious. His latest “pearls” were that at fault of high Radeon prices this generation is the “pricing politics of Nvidia” and that for “CUDA” you are fine with way weaker card than 4090.…right, cause Huang actually sets AMD prices as well, and compute tasks via CUDA do not scale with more cores unlike games, so considering 4090 for productivity is not justified.

Absolute clown. I only saw this level of shilling aimed toward someone else than AMD (intel) from the Userbenchmark admin in the recent Zen4 controversy.

 

[Kaluan]

sorry, but meh. Would take 24c 2495x, if it was priced right, over “24c” 13900k, even if the latter is faster in single thread.

Well then, sorry to say friend but after this hybrid R9 79x0X3D fake debacle, I am even more convinced Granite Ridge (Ryzen 8000) will be hybrid chiplet (not big.LITTLE but fast.FASTER lol) as a baseline. 12th gen/Microsoft/Linux layed the ground work, now 7000X3D is tweaking it for it's needs so Ryzen 8000 will have the scheduling ecosystem.



Anyway, tough crowd when AMD seemingly prefects consumer chiplet and now stacking technology each gen, while their competition hasn't even gotten the former off the ground yet.

I'll hash it up to leaker syndrome.
We haven't even seen 3rd party data or prices and people insist on being disappointed.

Guess you can't always please both tech junkies AND consumers at the same time.

 

[In2Photos]

Anyway, tough crowd when AMD seemingly prefects consumer chiplet and now stacking technology each gen, while their competition hasn't even gotten the former off the ground yet.

I'll hash it up to leaker syndrome.
We haven't even seen 3rd party data or prices and people insist on being disappointed.

Guess you can't always please both tech junkies AND consumers at the same time.

Agreed. I find it interesting that people are saying AMD didn't "fix" the frequency issue when using vcache. "Fix" implies that is was broken to start with. It's merely a hardware limitation of something in it's infancy. And let's get mad at AMD because the rumors were wrong, after all they were the ones that started all the rumors right? Rumors should just be considered a "wish list".

 

[Timmah!]

Agreed. I find it interesting that people are saying AMD didn't "fix" the frequency issue when using vcache. "Fix" implies that is was broken to start with. It's merely a hardware limitation of something in it's infancy. And let's get mad at AMD because the rumors were wrong, after all they were the ones that started all the rumors right? Rumors should just be considered a "wish list".

does not matter how you call it, fix or not, its semantics. So they did not fail to fix it, they failed to bring it out of its infancy. Either way, its clearly not a feature.

personally, i mostly wonder how can it be a voltage issue, if the cpu can have different voltages for its different parts - i am pretty sure my skylake-x had separate voltages for core and uncore. I understood it was a problem with 5000 series, as there were maybe not pins on AM4 socket for this separate “voltage plane”, as vcache was not a thing, when AM4 was new, so they did not anticipate it, but AM5?

 

[moinmoin]

7900X3D and 7950X3D turned out to be slightly awkward "have the cake and eat it too" packages, with the V-Cache CCD still being limited in frequency but now accompanied by an unhindered top end CCD.

So theoretically this product allows consumers to avoid making any tradeoff each separate choice would introduce.

In practice this hybrid design leads to plenty questions about feasibility with the scheduler used etc. Maybe a simple manual selector would be good to have.

The worst downside I see in this approach is that V-Cache allows the use of slower RAM but now doing so will hit the other CCD.

 

[Racan]

The worst downside I see in this approach is that V-Cache allows the use of slower RAM but now doing so will hit the other CCD.

With DDR5 6000Mhz being the sweet spot for Zen 4 I don't think this is a big issue if you can also afford to buy 7900X3D and 7950X3D cpus.

 

[In2Photos]

7900X3D and 7950X3D turned out to be slightly awkward "have the cake and eat it too" packages, with the V-Cache CCD still being limited in frequency but now accompanied by an unhindered top end CCD.

So theoretically this product allows consumers to avoid making any tradeoff each separate choice would introduce.

In practice this hybrid design leads to plenty questions about feasibility with the scheduler used etc. Maybe a simple manual selector would be good to have.

The worst downside I see in this approach is that V-Cache allows the use of slower RAM but now doing so will hit the other CCD.

I don't think they are awkward, instead I feel like they provided the consumer with enough skus for you to make a choice based on the following scenarios.

Gaming only - 7800X3D
Productivity only - 7900X or 7950X
Mixed use - 7900X3D or 7950X3D

The only pushback seems to be about the core clocks being dialed back and lack of vcache on both CCDs. But where does a 7900X3D or 7950X3D with the same core clocks as the non X3D versions and double vcache really shine? What type of workloads? Seems like it would be very few use cases.

 

[Dave3000]

I wonder how the Windows 10, 11, and Linux thread schedulers will handle this for programs that favor larger L3 cache more than a little higher frequency and programs that favor a little higher frequency than more L3 cache or is it something that will have to be added to the scheduler and is if so, is it even possible?

 

[Dave3000]

I don't think they are awkward, instead I feel like they provided the consumer with enough skus for you to make a choice based on the following scenarios.

Gaming only - 7800X3D
Productivity only - 7900X or 7950X
Mixed use - 7900X3D or 7950X3D

The only pushback seems to be about the core clocks being dialed back and lack of vcache on both CCDs. But where does a 7900X3D or 7950X3D with the same core clocks as the non X3D versions and double vcache really shine? What type of workloads? Seems like it would be very few use cases.

I guess you would have to buy the appropriate CPU for what you are going to do on your system, like you listed. For pure productivity and not gaming, I think the 7950X would be best out of all those you listed. For gaming I think the 7950X3D would offer the best flexibility including gaming flexibility, if you use a 3rd party program to manually assign which CCD the game will use and that's if the 3rd party program will support those CPU's. However, I'm not sure if it will be recommended to use a 3rd party scheduler such as Process Lasso for optimal application performance of the 7900X3D and 7950X3D or if Windows will be able to recognized what the application favors (v-cache/lower frequency or no v-cache/higher frequency) and assign the optimal CCD to the application accordingly. I wonder how consistent the the 7900X3D and 7950X3D will be in software that uses over 8 cores since each CCD has different performance characteristics on those CPU's. The reviews aren't out yet so I'm only presuming.

 

[Mopetar]

AMD had a good thing going with Zen2 and Zen3. Why did they decide to screw up their huge power consumption advantage with Zen4?

After looking at the AT article on power scaling Zen 4 really falls off after 105W. My only guess is that they want to ensure greater headroom on AM5 for future products. Eventually we'll get more than 16C parts and not having the extra power would constrain those parts.

The board partners would be pretty upset if they're making boards that have to support TDPs for CPUs that aren't rated for them, so I'm guessing AMD decided to just push the chips beyond their efficiency point so board partners wouldn't complain.

What stops AMD from simply remapping vcache at the system level? Instead of throwing 32MB of L3 evenly across 16 cores, what stops them from adjusting it on the fly? If they wanted to throw 32MB at core0, and withhold the rest from core1--> core15, that could really mess with benchmarks. Likewise, a core having little benefit from L3 shouldn't hold any hostage from the others.

If it's a single core benchmark it will eventually get as much of the L3 cache as it can use since it's fully shared across all cores. There's not really a need (or a point) to dynamically adjust it like this. Never mind that doing this might require increasing the access time by a cycle or two to make sure the management policy is being followed.

Any core that's running a workload that fits inside its own L2 cache won't be evicting cache lines to the L3 very often, so it's not going to be taking up resources it doesn't really need.

 

[In2Photos]

Flight Simulator 2020 benefits greatly from the v-cache. There's a big difference in this game between the 5800X and 5800X3D.

But does it benefit from more than 8 cores? If not, the 7800X3D is still the best option, no?

 

[PJVol]

but now accompanied by an unhindered top end CCD.

Is it known for certain, or there's a reason why it can't be the other way around?

 

[Dave3000]

But does it benefit from more than 8 cores? If not, the 7800X3D is still the best option, no?

Based on the reviews I read, it doesn't seem so, at least for pure frame rate. However, I'm not sure if load times in this game benefit from more than 8 cores (CPU does play a role in load times). I think that 8 cores is enough for gaming for at least this console generation.

 

[moinmoin]

I don't think they are awkward, instead I feel like they provided the consumer with enough skus for you to make a choice based on the following scenarios.

They are awkward in the sense that they offer a selling point for which to make use of no OS and software is prepared for yet.

Is it known for certain, or there's a reason why it can't be the other way around?

If it could be the other way around 7800X3D would boast a significantly higher max frequency.

 

[PJVol]

If it could be the other way around 7800X3D would boast a significantly higher max frequency.

Sorry, I don't see how 7800x3d boost frequency is relevant to the 2ccd chips L3 layout?
You're assuming the top CCD is 3D stacked, and I asked why it's not the lower binned CCD.

 

[biostud]

Do you think that all core boost speed on the 3D cache models will be significantly slower than the regular versions?

 

[moinmoin]

Sorry, I don't get, what 7800x3d has to do with the 2ccd l3 layout?

We have 7700X at max 5.4GHz whereas the CCDs with V-Cache all appear to max out at 5.0GHz. That's what 7800X3D maxs out at, if more were possible with V-Cache surely 7800X3D would have seen that. 7900X3D appears to add a 2nd CCD without V-Cache maxing out at 5.6GHz, and 7950X3D at 5.7GHz, both matching the max frequency of 7900X and 7950X respectively.

 

[PJVol]

7900X3D appears to add a 2nd CCD without V-Cache

Ah... i just wrongly assumed you mean the top binned CCD has V-cache.

 

[nicalandia]

AMD Bergamo 10 Chiplet layout.

 

[jpiniero]

That's a lot of power being drawn by the IO dies.