Speculation: Ryzen 4000 series/Zen 3

[JoeRambo]

There's nothing magic about dynamically sharing resources as needed. Point is, today's CPUs are more and more optimized for extreme corner cases, leading to bountiful resources that are otherwise unused. It's a classical case of overprovisioning, and SMT is a valid way to make better use of all those "overprovided" resources.

AMD/Intel/IMB already have SMT, there are no questions about it being useful. But they don't have SMT64 even if that would help in some artificial corner case. So there are certainly variuos design time and production/performance costs attached to SMT implementation, even if they are not "magic". Is AMD Ryzen as wide as Power9? Not really.

 

[moinmoin]

AMD/Intel/IMB already have SMT, there are no questions about it being useful. But they don't have SMT64 even if that would help in some artificial corner case. So there are certainly variuos design time and production/performance costs attached to SMT implementation, even if they are not "magic". Is AMD Ryzen as wide as Power9? Not really.

You may not believe it, but AMD only introduced SMT2 with Zen two years ago, and that already was more efficient (and also more secure) than Intel's HT which means there were already enough shareable resources. Zen 2 increased that once more, and @looncraz noted that its width is comparable to Power 7 which does SMT4. We are in the speculation thread for Ryzen 4k (and I'm expecting it to use Zen 3) so we talk about which low hanging fruits there are for AMD to pick for the next round. To me evidence points to SMT4 possibly being one of those low hanging fruits.

Whether SMT4 then is actually pushed or disabled on the mainstream desktop is a whole different discussion. Feel free to propose some other low hanging fruits you expect to be picked then, I'm all ear (and think we had too few of them so far to be on page 5 already).

 

[Atari2600]

There's nothing magic about dynamically sharing resources as needed. Point is, today's CPUs are more and more optimized for extreme corner cases, leading to bountiful resources that are otherwise unused. It's a classical case of overprovisioning, and SMT is a valid way to make better use of all those "overprovided" resources.

What is the average benefit from SMT on vs SMT off?

 

[teejee]

Your tendency for load of bs claims is just incredible. Basically what You claim is: Intel has some sort of magic accelerator implemented in Ice Lake for LZMA, that GeekBench is somehow able to immediately implement and utilize ( I guess for hardcore morons it is additional proof of conspiracy and collusion between Intel and devs of Geekbench ).

Intel could use macro-op fusion in order to use new instructions also with older code. This is a well known concept already in use.
This requires that it is easy to identify the code patterns to be replaced by the new instructions though.

 

[Andrei.]

This is a pretty much useless IPC test because they only tested Specfp and I'm not even sure if this is a good comparison because they tested 8C vs 12C SKUs. A frequency normalisation is a bad decision as well, they have to clock all CPUs at the same frequency because they can't guarantee a perfect clock scaling, this is a really flawed IPC comparison from AT. By the way in Geekbench the IPC is exactly the same as well if both are using the same RAM. As you can see the IPC test from Stilt is by far the best.

We tested the whole int and fp from 2006 and 2017 so I don't know what you're talking about. The core count is also completely irrelevant.

And frequency normalisation is the only valid way of testing microarchitectural IPC beyond doing PMCs, fixing clocks will get you nowhere in terms of determining what's happening on the core side.

 

[mikk]

We tested the whole int and fp from 2006 and 2017 so I don't know what you're talking about. The core count is also completely irrelevant.

Exactly this is the problem, you only tested SPEC and nothing else.

And frequency normalisation is the only valid way of testing microarchitectural IPC beyond doing PMCs, fixing clocks will get you nowhere in terms of determining what's happening on the core side.

You have really no clue how to do this properly, no wonder. Also your gaming tests are crap as usual from Anandtech because they are bottlenecked from the GPU. Gaming tests in CPU reviews from anandtech are one of the worst in the net, this is a usual anandtech issue, they really can't do proper CPU gaming tests, this is so pointless.

 

[Abwx]

Exactly this is the problem, you only tested SPEC and nothing else.

You said that they ony tested SpecFP, he s on point since they tested with Spec_int 2006/2017 and Spec_FP 2006/2017.

Besides you once said that Cinebench was a best case for AMD despite this bench being designed exclusively with Intel software suite including ICC, so it s foremost opimized for Intel before anything else.

But if that s not enough you can check other renderers, Povray, Blender, Vray, they are all best cases as well for AMD and yield about the same numbers as Cinebench :

https://www.computerbase.de/2019-07/amd-ryzen-3000-test/3/#diagramm-test-pov-ray-multi

The only renderers wich put Intel ahead is Corona since it use Intel designed renderer Embree, so if i understand this should be the only one used for comparison..?.

You can check the numbers for Vray at THG, same story as above, and most remarkable at Comuterbase is the AVX256 test wich put the 3700X 11% faster than the 9900K.

Actually you took The Stilt first tests at face value and went to think that the IPC delta between Zen 1 and SKL was more than a few %.

The 15% improvement of Zen 2 allow to leapfrog Intel s best, Zen 2 has 6-10% better ST IPC and as much as 10-15% in MT.

 

[happy medium]

The 4000 series will be a refresh adding at best 15% more performance overall with slightly lower power usage.

 

[maddie]

The 4000 series will be a refresh adding at best 15% more performance overall with slightly lower power usage.

This got a laugh.

Only a 15% at best generational improvement folks. That's all, move along now.

 

[Thunder 57]

This got a laugh.

Only a 15% at best generational improvement folks. That's all, move along now.

Replace Zen 3 with Sunny Cove and watch how that number changes...

 

[Vattila]

By the way, regarding speculating about Zen 3, Milan and its chiplet and interconnect topology, we had some discussion in the 64-core EPYC Rome thread.

Here is my speculation from that thread about how Milan may look, assuming AMD tries to cram sixteen 8-core CPU chiplets onto the package using an active interposer housing the interconnect (network-on-chip) and all the I/O. The lighter green is the interposer, while the cyan is the speculated L4, also assumed to be housed in the active interposer. See the linked discussion thread above for more detail.

PS. With SMT-4 this monster would have 128 cores and 512 threads!

Of course, this is a rather crazy idea, but intuitively I think some use of interposer to reduce energy consumed by the interconnect makes sense.

 

[tamz_msc]

Exactly this is the problem, you only tested SPEC and nothing else.

And that is a problem because? As expected you are utterly clueless about how SPEC works.

You have really no clue how to do this properly, no wonder. Also your gaming tests are crap as usual from Anandtech because they are bottlenecked from the GPU. Gaming tests in CPU reviews from anandtech are one of the worst in the net, this is a usual anandtech issue, they really can't do proper CPU gaming tests, this is so pointless.

You now claim to be more knowledgeable than Andrei. That is a big joke. Also you are deflecting the argument by bringing in gaming benchmarks because from the looks of it you have nothing to argue about.

 

[Carfax83]

I actually have to agree with mikk about the gaming tests. I don't know why they used a GTX 1080, which is basically a midrange GPU these days. And not only that, the settings they used made some of the games GPU limited rather than CPU limited, which defeats the purpose of the entire review.

Honestly, the game lineup needs to be completely overhauled. They need more modern titles with greater diversity akin to what Computerbase.de has.

 

[happy medium]

This got a laugh.

Only a 15% at best generational improvement folks. That's all, move along now.

Zen 2000 series was about 11% better then Zen 1000 series.3000 series is about 15% better than the 2000 series. 8 cores vs 8 cores. So suddenly the 7nm +,16 core 4000 series we will be more than 15% better than the 7nm 16 core 3000 series?
Why would you think that?

 

[beginner99]

DDR5 memory.

I suspect that will be too early on consumer side and after all this is a feature of the IO die which is the beauty of the design. Server gets ddr5 and consumer can remain on cheaper RAM and possibly same socket.

Personally I feel that overdoing socket compatibility will only hinder advacnment as it limits a design from the start. But I don't expect for Ryzen4000 to change much. We will get same 16-cores max on AM4 with DDR4 possibly even with same IO die. Bit higher clocks and IPC. Nothing game-changing. I don't see why they should increase core count again on AM4.

Then ryzen5000 should break compatibility with new socket, ddr5, more cores, igpu on IO die, somewhere mid 2022.

 

[ksec]

I only read the first page and it is immediately clear people are talking about Ryzen 4000 CPU, and not APU, or Ryzen CPU with Zen 3 Core.

Which is different to Ryzen 4000 APU as that will likely be using Zen 2. ( Unless they name the next Gen Ryzen APU into something else )

I wonder why OP didn't just ask about changes / speculation in Zen 3 instead.

 

[Abwx]

Zen 2000 series was about 11% better then Zen 1000 series.3000 series is about 15% better than the 2000 series. 8 cores vs 8 cores. So suddenly the 7nm +,16 core 4000 series we will be more than 15% better than the 7nm 16 core 3000 series?
Why would you think that?

13% for 2700X vs 1800X and 17% for 65W 3700X vs 105W 2700X, so at equal TDP the 3800X should be roughly 20% better than the 2700X :

https://www.computerbase.de/2019-07...performancerating-fuer-anwendungen-multi-core

Even it s linked on this page you still had to pull random numbers out of nowhere.

Anyway thoses actual numbers show that from the 1800X to the 3800X there should be more than 35% improvement, we can see that the 3600 (65W..) match the 1800X despite the latter having 33% more cores and 50% higher TDP, all this while Intel kept the same design.

 

[maddie]

Zen 2000 series was about 11% better then Zen 1000 series.3000 series is about 15% better than the 2000 series. 8 cores vs 8 cores. So suddenly the 7nm +,16 core 4000 series we will be more than 15% better than the 7nm 16 core 3000 series?
Why would you think that?

You're missing the sarcasm.

Who would not be happy with these IPC increases? Even using your numbers of 11%, 15% and 15% in three consecutive years we get a (1.1*1.15*1.15) = 1.45 or a 45% increase. Do you realize that is almost as big as the jump from Vishera to the original Zen? That means an original Zen owner (sorry guys) will next year feel as an 8350 owner did when Zen1 arrived.

 

[birdie]

 

[ksec]

I am guessing Zen 3 gets ~15% IPC improvement, that would be inline with Icelake IPC, and 7nm+. So along with Clock Speed improvement it should be close to 20% increase in performance.

Zen 4 will be Zen 3 in 5nm, big changes to I/O Die, DDR 5 and new Socket. Basically the Zen 3 Performance but much lower power usage.

 

[Abwx]

I am guessing Zen 3 gets ~15% IPC improvement, that would be inline with Icelake IPC, and 7nm+. So along with Clock Speed improvement it should be close to 20% increase in performance.

.

With 15% gain they will keep the lead they have currently, as said ad nauseam Zen 2 has quite better IPC than SKL, look like people are still not accustomed to the new realities, so here a reminder to help :

https://www.computerbase.de/2019-07...performancerating-fuer-anwendungen-multi-core

4.1GHz all core turbo 3700X match a 4.7GHz all cores turbo 9900K, how much better IPC does this make ?.

 

[Elfear]

I am guessing Zen 3 gets ~15% IPC improvement, that would be inline with Icelake IPC, and 7nm+. So along with Clock Speed improvement it should be close to 20% increase in performance.

Zen 4 will be Zen 3 in 5nm, big changes to I/O Die, DDR 5 and new Socket. Basically the Zen 3 Performance but much lower power usage.

I would love to see another 15% IPC increase but I seriously doubt it for Zen 3. I think we'll see somewhere around 5-10% with maybe a 5% increase in clocks.

 

[HurleyBird]

4.1GHz all core turbo 3700X match a 4.7GHz all cores turbo 9900K, how much better IPC does this make ?.

Zen 2 generally has better multi-core scaling, and yields way higher in SMT.

And this is something that is really starting to annoy me, because the term "IPC" has become far too fuzzy now. There's ST IPC, MT IPC, and MT w/SMT IPC, and they all paint a different picture. On the one end, you can say that Zen 2 just nudges past Skylake's IPC, as demonstrated in The Stilt's test suite. And on the other end, you can say that Zen 2 is blasting past Skylake by ~14-15% as shown in your computerbase.de link.

Which is the more correct number? Well, there's a lot of subjectivity no matter what you do, but for the perspective of forecasting the performance of future architectures I'd say that depends on where the bottlenecks are. Eg. How much of Zen 2's relative multi-threaded uplift versus Skylake is due to things that make Zen 2 slower at single threaded tasks verses things that make Skylake slower at multi-threaded?

Eg. Let's say that the entirety of Zen 2's MT advantage is explained by heavily multi-threaded tasks being less latency sensitive. As soon as AMD fixes their latency issues, which are artificially imposed not from architecture but from chiplets and interconnect, then single threaded IPC shoots way up. In that case it would be most correct to say that Zen 2 indeed has ~14-15% higher IPC than Sklyake, and vice versa if it isn't. The truth is probably in the middle. But where?

So where does that leave us? People can pick and chose the numbers that make their arguments look good, and they'll go on doing that. But in terms of gauging Zen 2's architectural advantage over Skylake, I think you can place a reasonably educated guess between 5-10% as a fair middle ground.

But also note that mitigations make these murky waters even more opaque. If mitigations can be reasonably fixed in hardware, or otherwise worked around in ways that either eliminate or significantly reduce their performance cost, then those shouldn't count against Intel for the purpose of forecasting future products.

 

[ondma]

The usual questions spring to mind:

1. Will it support currently existing motherboards (300/400/500 series chipsets)?
2. What kind of IPC increase are we talking about?
3. Will AMD manage to squeeze more frequencies?
4. What node will it use?
5. What will be its TDP?
6. Will it support AVX512 instructions?
7. When and if we can expect Ryzen 4000 CPUs with modern onboard graphics (e.g. Navi10/Navi20)?

More to the point, when will they work out the bugs of Zen 2?

 

[happy medium]

I would love to see another 15% IPC increase but I seriously doubt it for Zen 3. I think we'll see somewhere around 5-10% with maybe a 5% increase in clocks.

More or less what I was saying. I agree.