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!

 

[deasd]

I think the thing that is most to blame is AMD decide to push everything to limitation just for getting ~5% performance. yeah it is a disappointment that efficiency is no better than 5950X. But still this is not a trend that started by AMD so I wouldn't complain that much.
Also there's a lot of room that could be improved by users. Look at the TDP 105watts ECO mode, it is truly best level of silicon on earth!


edit: I saw so many guys claimed there would be 7900X3D, 7950X3D in the future, but I'm afraid this is just rumor.... or am I missing something?

 

[tamz_msc]

On that chart there are many choices better for the "average" person. The 5950x is less than $500 and the 7600x is $300 and both are better than the 12700k.

And you are talking about Sandybridge ??? Join the current decade dude.

How on earth is a 7600X better than a 12700K?

 

[Panino Manino]

People here already had time to think and have an opinion about this "intel Moment" by AMD?
To me it seems like AMD tried to squeeze a bit more performance to not risky being too much behind Rocket Lake (on some tasks and in gaming) with the new power limit and 95C temperature "target". Seeing how much energy and temperature drops with so little performance drop when reducing the TDP just a bit, it feels that AMD pushed Zen 4 a bit too far like Intel was/is doing.

Intel's extra little cores and better memory controller are making a difference to not let AMD reign uncontested.

 

[sammykhalifa]

Same here I don't give a fudge about temps but power consumption is a major factor me.

Laptops CPUs have been in the 95+ degrees for so long. It does not matter as long as there's no throttling. Right now efficiency will be king in the future it WILL dictate people's choice.

I used to say that, but I've been wondering why I can't keep my apartment cool in the Summer since WFH started. The PC running all day has to have something to do with that, right?

 

[HurleyBird]

AMD did not increase platform power limits (and cost) on AM5 to accommodate power hungry Zen 4 chips pushed past any reasonable limit for a meagre ~5% performance gain. Seems obvious that AM5 power delivery was bumped for the sake of potential future products that raise the core count and might need room to stretch their legs. Zen 4 only consumes as much power as it does because why shouldn't it when there's free performance on the table and 5% could make the difference between winning or losing to Raptor Lake, which will also be over-juiced? I do wish that AMD had defaulted to sane TDPs with the option to use higher TDPs for the extra ~5%, rather than the other way around, though.

 

[RnR_au]

BUT, when I get mine, I want to see what 105 ECO mode OR the -0.1 vcore trick, and see what happens to performance.

Kepler_L2 has a special request

https://twitter.com/x/status/1574547181729611777

 

[igor_kavinski]

There's no way you could make this remotely commercially viable.

I hope someone from AMD sees your post and takes up the challenge to prove you wrong

 

[igor_kavinski]

The French outlet Le Comptoir du Hardware tested X-Plane 11.

99 percentile:

View attachment 68185

Average:

View attachment 68186

Test • AMD Ryzen 9 7950X, 7900X, Ryzen 7 7700X & Ryzen 5 7600X : Zen 4, AM5, X670E, B650E & DDR5

Les Ryzen 9 7950X & Ryzen 7 7700X ainsi que leur chipset compagnon X670E passent à la moulinette : comparaison avec les précédentes générations de processeurs sur des jeux, encodage, traitement d'image et vidéo, rendu 3D, compression, encryptage, compilation et divers benchmarks synthétiques et...

www.comptoir-hardware.com

So Intel has the edge in X-Plane 11 in percentile frame rates at least for the 12900K. Keep in mind that these tests were done with the official supported memory frequency on both platforms - 4800 MT/s for Alder Lake and 5200 MT/s for Zen 4.

7700X@65W looks like a pretty good choice for the average X-planer.

 

[BorisTheBlade82]

I need more of this kind of charts:

@BorisTheBlade82 do you plan to run PES on a Zen 4 chip anytime soon?

Andreas Schilling from Hardwareluxx already sent me some results 😉
I will try to incorporate them this afternoon.

 

[zir_blazer]

I didn't read this Thread last few pages, but... Did anyone already ranted about the absurd bottleneck than the X670/X670E Chipset is where you have TWO daisy chained chiplets that depends on a single PCIe 4.0 4x upstream link? Not even PCIe 5.0 (Which according to the AnandTech review, the Processors actually supports for the Chipset lanes but the current Chipsets do not), just 4.0. A single PCIe 4.0 NVMe SSD and you're pretty much saturating it, add it a ton of USB and SATA devices. It is way too much I/O multiplexing for such a mediocre uplink.
Given the fact than AM5 now has 28 total PCIe lanes coming from the Processor, they could have used two chiplets with their own separate uplinks (Sort of like ThreadRipper Pro having a 8x uplink instead of 4x) and still provide 4 PCIe 5.0 lanes for a single NVMe drive, but no, better to make a giant bottleneck.

WHY I'm always dissapointed by AMD at the platform level. WHY.

 

[igor_kavinski]

WHY I'm always dissapointed by AMD at the platform level. WHY.

Coz their core strength is CPUs.

 

[ondma]

I used to say that, but I've been wondering why I can't keep my apartment cool in the Summer since WFH started. The PC running all day has to have something to do with that, right?

Well, there is a difference between temperature and heat. Unless you are running a near max workload on a highly multithreaded program even a 100 or 200 watt cpu is not going to heat up an apartment that much. Most work from home would not utilize that much cpu power in any case. If one is gaming with a high power cpu and gpu (probably 2x the cpu power usage at least) you could be using 500 to 700 watts, which is getting close to the low setting on a space heater.

 

[coercitiv]

Did anyone already ranted about the absurd bottleneck than the X670/X670E Chipset is where you have TWO daisy chained chiplets that depends on a single PCIe 4.0 4x upstream link? Not even PCIe 5.0 (Which according to the AnandTech review, the Processors actually supports for the Chipset lanes but the current Chipsets do not), just 4.0. A single PCIe 4.0 NVMe SSD and you're pretty much saturating it, add it a ton of USB and SATA devices. It is way too much I/O multiplexing for such a mediocre uplink.

Some people complained and I had to volunteer as devil's advocate. Let's recap:



On AM5 the CPU alone offers the following direct connectivity:

• 16 PCIe 5.0 lanes for the GPU slot
• 4 PCIe 5.0 lanes for an M2 drive
• 4 PCIe 5.0 lanes for another M2 drive or Thunderbolt 4 / USB4
• 4 USB 3.2 lanes
• 1 USB 2.0 lane

So AMD did the hard work even before adding the chipset connection, since the user can either connect 2x NVME PCIe 5.0 drives or 1x NVME PCIe 5.0 drive + 2x PCIe 4.0 drives hosted on an extension card. All one has to do is choose the right board for the job, since the second bundle of general purpose PCIe lanes will be used differently from one board model to another.

The chipset's purpose, or at least my interpretation in the circumstances, is to provide the needed amount of I/O that is comparatively much slower than PCIe 4.0, all those extra USB ports that may simply get filled with low bandwidth dongles and devices, some SATA drives or PCIe 3.0 drives meant for secondary storage, additional LAN or WiFi, a sound card or a capture card etc.

The only problem I can see with this approach is the user needs to plan in advance, they need to understand the strengths and weaknesses of the design and buy the motherboard configuration that fits their needs. The upside is you can get a lot more very fast storage connected straight to the CPU, and that includes some external mass storage device connected via Thunderbolt 4 / USB 4 while completely bypassing the chipset.

 

[naad]

OK, then pic a wattage number, and who wins most if not all benchmarks ? I am pretty sure thats Zen 4. Raptor lake may be better, but based on what I have seen, even more power used (more cores) and marginal performance increase.

Now to sum up (and this is pure opinion) Intel created the E-cores to save power, but kept the power hungry P-cores to help performance. But their node was inferior (and still is) and that is killing them in perf/watt. So they tweaked ADL, and added more e-cores to try and catch Zen 4, and it does not look like that is going to happen. Once they get on a comparable node to AMD, things may get interesting. Thats going to be a while IMO.

Edit, and 100 watts more is not 10% and they will still lose from the leaks.

E-cores or "cinnebench accelerators" as I've recently started calling them only exist because the P-cores are extremely area and power inefficient. Intel wouldn't even be making this clumsy hybrid design that needs a constant updates to software crutch (thread director) and disabling of half the ISA extensions if their P-cores weren't literally double the size of their 2 year old competitor while being around 30% less efficient. There's a reason there's no hybrid nonsense in servers and only specialized workloads use E-core only SKU.

E-core spam will continue until Intel can get a proper area and power P-core out, that's probably not until 2025.

 

[poke01]

Wow seems like like people in thread lean hard to AMD. Oh well, the chart shows that the 12700K is better in handbrake than a 7600X.
The 5950x is on a dead platform, the 12700K is not as it can be upgraded to 13th gen.

Is it so hard for you to accept that Intel has a good value chip?

Also I brought up Sandy bridge as I was talking about market leader and Intel was clearly the market leader then but you probably think differently and believe otherwise. That won't surprise me.

On that chart there are many choices better for the "average" person. The 5950x is less than $500 and the 7600x is $300 and both are better than the 12700k.

And you are talking about Sandybridge ??? Join the current decade dude.

You want to keep talking about the 12700K?
We have a thread for that.
But its not this Zen4 spec thread.


esquared
Anandtech Forum Director

 

[deasd]

OK......MSI internal materials that knocking my head quite hard again:




limiting temperature to just 65C, only 5% performance lost.

DIY cooler, no perf lost.........DAFAK

 

[igor_kavinski]

Oh well, the chart shows that the 12700K is better in handbrake than a 7600X.

12700K certainly has its merits. Some applications love Intel architecture and there's the possibility of going very high with DDR5 speeds on Intel's side while AMD may need to work out some kinks in their IMC before it is as mature as Intel's.

One thing I don't understand. AMD got Zen3+ working with LPDDR5-6400 but Zen 4 sweet spot is a comparatively paltry 6000 MT/s? There really must have been two different teams, with the Zen 4 team not having had that much time to play with DDR5.

 

[Timorous]

The French outlet Le Comptoir du Hardware tested X-Plane 11.

99 percentile:

View attachment 68185

Average:

View attachment 68186

Test • AMD Ryzen 9 7950X, 7900X, Ryzen 7 7700X & Ryzen 5 7600X : Zen 4, AM5, X670E, B650E & DDR5

Les Ryzen 9 7950X & Ryzen 7 7700X ainsi que leur chipset compagnon X670E passent à la moulinette : comparaison avec les précédentes générations de processeurs sur des jeux, encodage, traitement d'image et vidéo, rendu 3D, compression, encryptage, compilation et divers benchmarks synthétiques et...

www.comptoir-hardware.com

So Intel has the edge in X-Plane 11 in percentile frame rates at least for the 12900K. Keep in mind that these tests were done with the official supported memory frequency on both platforms - 4800 MT/s for Alder Lake and 5200 MT/s for Zen 4.

Interesting that the lower power limit on the 7950 actually increased the minimums by 5 fps (about 7%).

EDIT:

Just quickly flicked through the review and the lower powerlimit seems to offer no change or an improvement for the most part. There are a couple of outliers.

I get the feeling this range of CPUs has a lot of tweaking to unlock and once there are more tools to help with that I can see some pretty reasonable gains to be made.

Out of the box they seem okay. Better than what was 1st presented when AMD said >15% ST performance gains and >5ghz. What I do find impressive is the almost linear scaling going from the single CCD to the dual CCD parts. Something that was not happening prior due to the low TDP limits of Zen 2 and Zen 3 dual chiplet skus.

Gaming seems to be a mixed bag. The 5800X3D does better on average than I thought it would. I expected Zen 4 to be ahead by 10% +/- in a standard test suite like HUB use with some strong outliers in favour of the 3D like for the racing / flight sims, stellaris and anything else that can really use the cache. I thought there would be a bit of daylight between the 12900K and the Zen 4 SKUs but there really is not so I do expect RPL to take the gaming crown but I do wonder what power draw that will be at and it won't be much.

I expect the RPL parts to do well in productivity vs the 7600X and 7700X but I don't think the higher up parts are catching the 7900X and 7950X. The lower end ADL refresh 13th gen parts are likely to struggle so a future 7600 when DDR5 prices reduce and B series mobos are available could make for a good budget part ala the 13600 or 13400 because I expect it to have a reasonable performance advantage across the board.

Pricing will also matter if the 13600K does come in at around £350 vs the £300 for the 7600X and the £400 for the 7700X then they all occupy there own price slot where both are valid options depending on use case.

The Zen 4 3D parts are going to romp in gaming though. If AMD can reduce the clock deficit required by the 5800X3D and the rumour of a large than 64MB additional slice is true (or a 2hi stack!) then a 20-25% gain in gaming seems pretty easy to achieve and that will by far make it the gaming king.

 

[zir_blazer]

So AMD did the hard work even before adding the chipset connection, since the user can either connect 2x NVME PCIe 5.0 drives or 1x NVME PCIe 5.0 drive + 2x PCIe 4.0 drives hosted on an extension card. All one has to do is choose the right board for the job, since the second bundle of general purpose PCIe lanes will be used differently from one board model to another.

Using a PCIe 5.0 4x slot to get 2 PCIe 4.0 NVMe drives would require a PCIe Switch on such card and I can absolutely guarantee you than it will be too expensive to even consider it a valid option.

The chipset's purpose, or at least my interpretation in the circumstances, is to provide the needed amount of I/O that is comparatively much slower than PCIe 4.0, all those extra USB ports that may simply get filled with low bandwidth dongles and devices, some SATA drives or PCIe 3.0 drives meant for secondary storage, additional LAN or WiFi, a sound card or a capture card etc.

The only problem I can see with this approach is the user needs to plan in advance, they need to understand the strengths and weaknesses of the design and buy the motherboard configuration that fits their needs. The upside is you can get a lot more very fast storage connected straight to the CPU, and that includes some external mass storage device connected via Thunderbolt 4 / USB 4 while completely bypassing the chipset.

Modern Chipsets are pretty much I/O multiplexers because I/O is pin intensive and you would need a gargantuan socket to have all the connectivity coming from the Processor package itself. Point is, daisy chaining TWO Chipsets seems an awful idea when you could have used the new 4 lanes as another uplink. Also, consider than the downstream chiplet has a higher latency than if it was wired to a direct uplink.
I'm not expecting than anyone that actually needs that level of connectivity will not use several things simultaneously in such a way that the uplink gets bottlenecked. The B650 is fine, but the X670 as two daisy chained B650 is horrendous.

Also, while I didn't saw AM5 Motherboards specifications, I'm pretty sure than the most typical setup will be to see one 16x PCIe 5.0 Slot (Or two 8x/8x via bifurcation on high end boards), and two 4x PCIe 5.0 M.2 Slots for NVMe SSDs, but it will be rare to see one of those two ports in PCIe Slot form. So it means than most add in cards will be on the Chipset, and most likely you will need to go out of your way if you want a third Processor PCIe Slot. If the rest of the PCIe Slots are from the Chipset, more cards are adding a further bottleneck.

 

[inf64]

UB posted 7950X and 7900X officially. The results are way way lower than what we had in the leaked scores that were not official. I'll give them benefit of the doubt as we have only 1 sample for each, but this is super fishy.

Edit: lol they removed the old leaked scores, they cannot be found in the database any more... Pathetic.

 

[cellarnoise]

What is needed is a price war for us consumers. Something like a 31.457%.. across the motherboard reduction would be a good computational start.

I don't understand why a direct die / with comparable risk warranty option would be bad either. These are gold plated and should be treated as such

I know my Athlon 1200 was like gold back in the day with cork? CPU Side pads to balance out the direct die heatsink load?

Risk should have the opportunity of reward?

 

[Timorous]

So for the 3D parts I think a 30% boost may actually be doable. Especially for the 7800X3D.

ComputerBase IPC If you check this out at a fixed 4.4Ghz across a pretty comprehensive gaming suite the 5800X3D was 29% faster than the 5800X. Provided the 7800X3D manages to maintain similar clock speeds to the 7700X I don't see why the same won't apply to Zen 4. If the rumour of a large slab of cache is true then it might be even greater still so it looks pretty obvious to me that if you are gaming focused waiting for the 3D parts is probably the best move because I don't see RPL competing with that at all, there is going to be clear daylight between RPL and Zen 4 3D in gaming.

 

[DAPUNISHER]

Wow seems like like people in thread lean hard to AMD. Oh well, the chart shows that the 12700K is better in handbrake than a 7600X.
The 5950x is on a dead platform, the 12700K is not as it can be upgraded to 13th gen.

Is it so hard for you to accept that Intel has a good value chip?

Also I brought up Sandy bridge as I was talking about market leader and Intel was clearly the market leader then but you probably think differently and believe otherwise. That won't surprise me.

I'm not down voting you due to bias. I am down voting you because you keep pimping Intel in the Zen 4 thread. Save us the sales pitch if you would be so kind.

 

[DAPUNISHER]

After watching and reading a bunch of different reviews, a few things were of note to me.

Memory training can take a long time.
Mainboard makers are going to have adapter kits if you want to use a cooler that requires its own back plate.
DDR5 6400 can improve performance in some games.
Curve optimizer is money with Zen 4.
The meme lives on - AKA undervolting and overclocking AMD hardware is basically mandatory to get the best experience.
Reviewers are all over the place. One goober even used DDR5 5200 instead of the provided 6000 kit.
Tiny Tom is old school and having trouble accepting the idea that 95c is okay. He also had trouble overclocking where Wendell did not. I think Tom relied on the UEFI where as Wendell used the curve optimizer.

 

[coercitiv]

Modern Chipsets are pretty much I/O multiplexers because I/O is pin intensive and you would need a gargantuan socket to have all the connectivity coming from the Processor package itself. Point is, daisy chaining TWO Chipsets seems an awful idea when you could have used the new 4 lanes as another uplink.

And my point was AMD prioritized another direct PCIe 5.0 CPU link over a secondary independent chipset link. Whatever you intend to connect through the chipset that is bandwidth intensive, you link directly to the CPU instead. The only caveat is the user needs to plan a bit ahead, because the second bundle of general purpose lanes lanes may come in different ending formats (NVME slot, PCIe slot, onboard USB 4.0 etc).

This solution is not ideal, just a trade-off: it simplifies layout and lowers cost while forcing the user to commit to a certain utilization pattern. Your solution would increase BOM, even more so if at least one chipset link would be PCIe 5.0 so that we do not lose bandwidth. The AM5 board prices are already a nasty surprise...