AMD Zen 4 Ryzen 7000 Series Update: 8–10% IPC Uplift, 25% More Perf-Per-Watt, More than 35% Overall Performance Gain, 5.5GHz & V-Cache Chips Coming

[ZetZet]

Just now, KaitouX said:

It's DDR5 only, that alone makes it pretty much impossible to be priced well compared to Alder Lake.

And again due to DDR5 lower SKUs are probably going to be horrible value, plus motherboards are probably be similar priced or more expensive than B660 and Z690. Alder Lake also has a significant thread count advantage over AMD currently in the lower SKUs, which might increase with Raptor Lake.

Intel should be also releasing Raptor Lake around the same time as Zen 4, so Zen 4 failing to match Alder Lake is pretty bad when you consider that too.

DDR5 is already coming down in price close to where it barely matters, AMD mobos are always cheaper too and the leaks are already pointing out that they will be cheap even on launch. Raptor Lake seems to be delayed already and we still have nothing about it's improvement. If they only add small cores to it I don't see how it can be so magical. 

[KaitouX]

1 minute ago, ZetZet said:

DDR5 is already coming down in price close to where it barely matters, AMD mobos are always cheaper too and the leaks are already pointing out that they will be cheap even on launch. Raptor Lake seems to be delayed already and we still have nothing about it's improvement. If they only add small cores to it I don't see how it can be so magical. 

Decent DDR5(like the ones used on the tests by AMD) costs around 3 to 4x the price of good DDR4. And while the price is coming down, it probably will still be over double the price of DDR4 when it gets released. The leaks put the motherboards prices around the same as current Intel boards. 

 

Intel already shown that Raptor lake is going to have 8+16C/32T, and said it would have up to double digits improvements. It doesn't need to be magical, as it would be compatible with current motherboards from Intel and supports DDR4.

 

Pricing is going to be the most important factor about Zen 4, but I don't see AMD selling a 8-core Zen 4 CPU for under $300 on launch, and that is about what would be needed to be a good option in my opinion. Hopefully they do it, but I don't really expect them to.

[ZetZet]

5 minutes ago, KaitouX said:

Decent DDR5(like the ones used on the tests by AMD) costs around 3 to 4x the price of good DDR4. And while the price is coming down, it probably will still be over double the price of DDR4 when it gets released. The leaks put the motherboards prices around the same as current Intel boards. 

6400 CL32 kits cost around 3 times of what 3600 CL16 costs. If you step down to 6000 CL36 it's only 2 times and the performance will still be higher, considering even "slow" DDR5 is fine in most tasks. 

5 minutes ago, KaitouX said:

Intel already shown that Raptor lake is going to have 8+16C/32T, and said it would have up to double digits improvements. It doesn't need to be magical, as it would be compatible with current motherboards from Intel and supports DDR4.

It's not going to have that performance with DDR4, why do you think Intel is pushing manufacturers to not promote DDR4 on Z790?

 

 

Also we still have like half a year until the release and then another half a year before early adopter tax gets taken off. So DDR5 is not going to be a problem.

[LAwLz]

11 hours ago, leadeater said:

So actual AVX-512 is part of Zen4, that's actually really big news.

Now the question becomes, what do they define as "AVX-512" and which products will include it?

It might just be servers, or it might just be one of the subsets of AVX-512.

 

It would be great if it was the base AVX-512 and maybe some more extensions on the consumer platform, but I would not be surprised it if turns out to be some server exclusive feature.

 

 

1 hour ago, ZetZet said:

What are you even saying. It's MORE than 15%, always, on everything. In what universe is that disappointing? Also if you do the math it will be more. 8% IPC and the clock speeds are at least 10% higher, because they are saying OVER 5.5GHz boost.

No, that's not how it works.

AMD did not say "you will see at least 15% higher single threaded performance in all workloads". 

When a manufacturer says "X% faster", they are referring to the tests they have done. I am 100% sure that there will be cases where zen4 do not provide 15% or more single-threaded uplift. 

[ZetZet]

Just now, LAwLz said:

I am 100% sure that there will be cases where zen4 do not provide 15% or more single-threaded uplift. 

I don't see how that's even possible considering the clocks increased by about that much alone. 

[LAwLz]

Just now, ZetZet said:

I don't see how that's even possible considering the clocks increased by about that much alone. 

Because CPUs are complex pieces of hardware that behaves differently depending on which instructions you run on them. 

Clock speed does not scale linearly in all workloads, and IPC is just a generalized measurement that does not take into consideration a lot of edge cases.

Also, all Zen4 processors will not be the same spec. So far AMD seems to have mostly talked about their top of the line chip. We don't know if the lower end chips (which, let's be honest, are the ones that people actually buy) will see an equal gain in frequency.

 

Anyway, it remains to be seen how Zen4 turns out. I am a bit more optimistic after this presentation, but I think people need to remember that these are first party numbers from a company trying to sell you a product. You should be heavily sceptical of any claims they make because it is very likely that they are trying to fool you. I mean, just look at the bar graphs they used in the presentation, where >25% is not linear and appears to be more than a 100% increase. Same with the graph showing ">35%".

[LAwLz]

1 hour ago, ZetZet said:

Which doesn't exactly matter, does it? Even if it's a bit slower it can be priced to be extremely competitive. Unless you're only looking at the top SKU and maximum performance, which to me isn't the most important. If they can match Adler Lake and push the cost down while also pushing wattage down that's a huge win. 

You say this, but then goes on to say:

1 hour ago, ZetZet said:

They already absolutely smash Adler Lake in MT, we already saw that. I don't see why ST wouldn't be close. I think you're expecting Raptor Lake to be a huge improvement, but so far they haven't given us anything at all. 

Let's make one thing very clear. AMD only "smash Alder Lake in MT" if we look at the highest end SKUs. Alder Lake is faster than Ryzen 5000 for both single- and multi-threaded workloads in most price segments.

For example the i5-12600K which is about 10% cheaper than the 5800X beats it in Cinebench MT by about 12%. 

 

 

1 hour ago, ZetZet said:

They already absolutely smash Adler Lake in MT, we already saw that. I don't see why ST wouldn't be close. I think you're expecting Raptor Lake to be a huge improvement, but so far they haven't given us anything at all. 

Raptor Lake will most likely be a huge improvement for MT workloads, since it seems like they will double the amount of E-cores.

It would be great if zen4 becomes competitive with Alder Lake in single threaded performance, because if it doesn't then AMD risk being beat in both single threaded and multi threaded workloads.

[Ydfhlx]

22 minutes ago, ZetZet said:

I don't see how that's even possible considering the clocks increased by about that much alone

Because performance doesn't scale linearly with clocks. A 10% higher clock on same architecture usually results in around 5-7% higher performance. That's because there's other stuff to consider, like cache or memory.

 

I can assure you that if Zen 4 was faster AMD would certainly say this.

 

By the way, I get flashbacks from Bulldozer launch. Relatively low IPC improvement, but high clocks, multi core over single core. Let's just hope it doesn't end like it did that time.

[porina]

25 minutes ago, LAwLz said:

IPC is just a generalized measurement that does not take into consideration a lot of edge cases.

To expand on this, when AMD or Intel talk IPC, they use an average of a variety of workloads. Some will be lower, some will be higher.

[leadeater]

2 hours ago, LAwLz said:

Now the question becomes, what do they define as "AVX-512" and which products will include it?

It might just be servers, or it might just be one of the subsets of AVX-512.

 

It would be great if it was the base AVX-512 and maybe some more extensions on the consumer platform, but I would not be surprised it if turns out to be some server exclusive feature.

No idea, probably not the end of the world if it's not enabled on Ryzen. There isn't really enough memory bandwidth most likely to do anything useful with it.

 

AVX-512 is AVX-512, if you are calling out that then it's at a minimum AVX-512F. AVX-VNNI is a very specific exception for processors that do not feature AVX-512 but have AI/ML instruction set support.

 

Quote

A later AVX-VNNI extension adds VEX encodings of these instructions which can only operate on 128- or 256-bit vectors. AVX-VNNI is not part of the AVX-512 suite, it does not require AVX-512F and can be implemented independently.

 

So you don't say you have AVX-512 support if you have AVX-VNNI ("AI") because it's not part of AVX-512 at all. AVX512-VNNI != AVX-VNNI, yea confusing I know heh.

 

Also all the extra AVX-512 subset stuff, that's an Intel thing not an AVX-512 thing. Since AMD has never done AVX-512 before we can't know exactly what their AVX-512 instruction set actually covers other than the base minimum AVX-512F which is the core standard.

 

Edit:

Oops I forgot AMD gave an official answer to this information lol

Quote

Your presentation mentioned "AI acceleration." Is that AVX-512 or something more exotic, like Intel GNA?
Yes. Specifically, AVX 512 VNNI for neural networking and AVX 512 BLOAT16 for inferencing. Both are pretty nice speedups, we're not using a fixed-function acceleration, this could be something we could do with our Xilinx acquisition. We are starting to see more consumer applicability of AI workloads, like video upscaling, which has grown a lot in the last two years. I think there is a general trend for the average enthusiast to take on more AI-type workloads. The time felt right to bring these features into the chip, given we moved to a smaller process node with better performance, power, and area capabilities.

https://www.techpowerup.com/review/amd-zen-4-ryzen-7000-technical-details/

 

So interestingly potentially not AVX-512F at all 

 

Anyway sounds like going to be part of Ryzen.

[Ultraforce]

11 hours ago, cmndr said:

 

If you do shorthand you might see the Sigma symbol with just an n under it (go across al entries).

This pops up in statistics a lot, where you have an arbitrary (and potentially unknown up front) number of data points to go through.

 

The above is the formula for taking the average (Sigma means SUM up). so from the first entry to the nth entry (1, 2, 3, ... , n) take all of your data points (so x_i is the data point in the ith position, so the first data point is x₁, the second x₂, the nth is x_n) sum them up and divide by the number.

 

I will say while I know that if asked I would have guessed NT was like some sort of version if it was Cinebench nT then it would be a lot more obvious.

[Doobeedoo]

Overall it looks very good so far. What I am mainly curious is to see the 3D cache SKU though, how much better than current one is, that is the bar for me. I hope, they will release 3D cache SKU sooner this time.

[LAwLz]

1 hour ago, leadeater said:

AVX-512 is AVX-512, if you are calling out that then it's at a minimum AVX-512F. AVX-VNNI is a very specific exception for processors that do not feature AVX-512 but have AI/ML instruction set support.

-snip-

So you don't say you have AVX-512 support if you have AVX-VNNI ("AI") because it's not part of AVX-512 at all. AVX512-VNNI != AVX-VNNI, yea confusing I know heh.

1 hour ago, leadeater said:

Oops I forgot AMD gave an official answer to this information lol

 

-snip-

 

So interestingly potentially not AVX-512F at all 

 

This is why I will remain skeptical until zen4 is out and we have hard evidence of them supporting AVX-512F. As long as AMD is being somewhat vague and talking about AVX-512 in the same sentences as VNNI I will be hesitant to believe that they actually support AVX-512F.

 

"AVX-512" might not be "AVX-512" when it is talked about in marketing.

[Zodiark1593]

6 hours ago, DANK_AS_gay said:

Would Fermi 2.0 be a good candidate for this type of work?

Most likely not. 
 

With the Cell architecture, the PPE (main CPU) dispatches tasks to the SPEs, which are fast vector processors themselves, which can handle a variety of code including data for shaders. The results are returned to the PPE (as the SPEs cannot directly address main memory). The host CPU needs that data back to continue the emulation, which is where latency can be detrimental. In console emulation, nearly every emulated component is dependent on one another. 

 

This is before considering the overhead of translating the code to something that will run on the host hardware. Being branch-heavy and reliant on per-core performance, this is certainly not something a GPU can handle itself, requiring the CPU to translate the code beforehand. 
 

Modern GPUs are very good at chewing through massive amounts of data, very quickly. However, it takes time to move data back and forth across the PCI-e bus. For most workloads, this isn’t a big deal, as you’re probably compute bound anyway. They’re not well suited to highly bursty, latency sensitive workloads. 

[Quackers101]

it is nice to see the clock go up.

but against intel, would they have a few issues?

also the 3D cache that AMD has dealt with and any issues it could have or improve for zen 4?

then how intel did better in some workloads or idle per watt.

[spartaman64]

7 hours ago, porina said:

Let's see what they do. If it does appear in consumer CPUs, it makes them much more interesting to me again.

 

So far it has been a bit chicken and egg. Software devs might not look at it until there is sufficient installed base. AMD may be doing their classic of letting someone else do the hard work on adoption before jumping in. Intel, is doing an Intel. It didn't appear in Intel consumer desktop CPUs until Rocket Lake, then officially disappeared again with Alder Lake. On mobile side it has been included in previous two generations: Ice Lake and Tiger Lake. If AMD really do implement AVX-512 we could see more uptake of the instruction set.

 

As for what it does, note AVX-512 is a bit of an umbrella for a whole bunch of mandatory and optional stuff. The core part improves upon AVX2. If software can make good use of AVX2, it could potentially see decent gains from implementing AVX-512. For Prime95 like workloads I've seen +40% IPC on one unit implementations (as used in consumer tier CPUs) and +80% IPC on two unit implementations (as seen in Intel HEDT and some server versions).

 

A side problem also has been that since AVX-512 can do a LOT of work fast, it also uses a lot of power while doing so. Again for Prime95 like workloads, perf/W seems about constant relative to not using it, so more work is more power. In the past Intel have managed this by significantly dropping the clock when AVX-512 code runs. It seems like on newer CPUs this is better managed. Strongly using AVX-512 will still see benefits. If you only make light use of it, the clock drop might hurt non-AVX-512 work going through at the same time. AMD are not immune to this either if you look at their AVX2 behaviour. When running under a fixed power limit (stock) that too drops clocks more than other workloads which don't use it.

 

Without going back to check, wasn't the 15% claim for single thread? This 35% claim is multi-thread.

 

 

On the claims here, so we get >25% perf/w but >35% overall. Implication that last 10% is from higher power. Since scaling is always worse as you go up the curve, expect that power to go up more than that. Taking 142W PPT of 5950X: 142/1.25*1.35=153W as a minimum. Expect higher but it could be lower depending on what the >25%/>35% numbers actually are.

idk stuff that avx512 would be good for you'd probably want to do on the gpu anyways

[porina]

15 minutes ago, spartaman64 said:

idk stuff that avx512 would be good for you'd probably want to do on the gpu anyways

It's a classic case of use the right tool for the job. Neither CPUs or GPUs do everything well. For many tasks GPUs are too limited in scope compared to a CPU, especially the more limited consumer tier GPUs.

[leadeater]

2 hours ago, porina said:

It's a classic case of use the right tool for the job. Neither CPUs or GPUs do everything well. For many tasks GPUs are too limited in scope compared to a CPU, especially the more limited consumer tier GPUs.

Data movement also has huge penalties so doing lots of small tasks on the CPU twice as slow or more could be end to end faster than copy to GPU, processing, copying back to system memory,  processing returned data in CPU etc 

 

Without true unified memory and zero copy you really want to give GPUs big long jobs rather than small ones, or if you give it small ones batch them in to many and ensure you keep the GPU feed well.

[cmndr]

11 hours ago, Rauten said:

Now with Zen4, which is supposed to be a brand new architecture, it's like barely any improvement has been achieved. Most of the gains could easily be put down to the speed increase and the node reduction.

Being able to scale to higher clock speeds, without losing IPC, IS an architectural improvement in all likelihood.
I wouldn't be surprised if AMD lengthened their pipeline a dash (hurts IPC) and then made improvements elsewhere to compensate.

There is no free lunch in CPU design. Most improvements in one area hurt other areas (want more IPC? that usually means lower clocks or higher die costs or more power draw).
https://www.lighterra.com/papers/modernmicroprocessors/

 

12 hours ago, KaitouX said:

Probably because unless it's 25%+ average it doesn't even beat Alder Lake. Also, the >15% comes from Cinebench R23 ST alone, not from multiple workloads according to the footnotes, it's at least 15% improvement in Cinebench R23 ST, which is well below the 12900K ~23% or the 12900KS ~30% improvement over Zen 3, other workloads can be below that, even though I would hope CB ST is in the lower side.

Zen 3 currently ties or slightly edges out ADL at MT while using substantially less power in most cases.
It could very well be the case that Zen 3 wouldn't beat RaptorLake though.

It's very probable that Zen 4 (at least without 3dVcache) won't have the ST crown vs RaptorLake but WOULD have a modest MT crown.

[cmndr]

12 hours ago, LAwLz said:

You say this, but then goes on to say:

Let's make one thing very clear. AMD only "smash Alder Lake in MT" if we look at the highest end SKUs. Alder Lake is faster than Ryzen 5000 for both single- and multi-threaded workloads in most price segments.

For example the i5-12600K which is about 10% cheaper than the 5800X beats it in Cinebench MT by about 12%. 

 

 

Raptor Lake will most likely be a huge improvement for MT workloads, since it seems like they will double the amount of E-cores.

It would be great if zen4 becomes competitive with Alder Lake in single threaded performance, because if it doesn't then AMD risk being beat in both single threaded and multi threaded workloads.

So all true but let's keep things in context.

At the moment a 12700k and a 5900x are very similar in MT. Modest edge to the 5900x. Similar story for 12900 and 5950x. ADL is obviously out front in ST. Let's say the 5950x is 10% ahead.

In the case of the 12900k, the E cores contribute about 1/3rd of the MT performance.

Going to the other end - if the 13900k has 2x the E cores, then you'd naively expect a 33% boost there... plus maybe another 10% uplift from the P cores getting buffed. This could certainly mean a 40-50% gain in MT.

 

Looking at the 7950x - if you bump it up by 35%, then you'd expect around parity with Raptor Lake on MT. The clocks and IPC on Zen4 also appear to be similar to Alderlake.
At that point you have to wonder how much of a boost to ST Raptorlake gets. 10% is my guess. How does that compare to any gain for 3dvcache?
 

This really is an instance where use case will matter.

[leadeater]

18 minutes ago, cmndr said:

This really is an instance where use case will matter.

Lets be honest, everyone here mostly only cares about gaming performance and if all you want is the top end then a 5800X3D and 12900K are functionally equivalent for this with the 5800X3D being cheaper and more efficient. That said right now today if you were buying something, not waiting, the 12900K is the better choice since the platform it lives on has an upgrade path where anything Zen3 does not.

 

Zen 4 with 3D V-Cache will likely live in the same situation we have now for gaming against Raptor Lake, not sure who will "win" but it won't "matter".

[cmndr]

36 minutes ago, leadeater said:

Lets be honest, everyone here mostly only cares about gaming performance and if all you want is the top end then a 5800X3D and 12900K are functionally equivalent for this with the 5800X3D being cheaper and more efficient. That said right now today if you were buying something, not waiting, the 12900K is the better choice since the platform it lives on has an upgrade path where anything Zen3 does not.

 

Zen 4 with 3D V-Cache will likely live in the same situation we have now for gaming against Raptor Lake, not sure who will "win" but it won't "matter".

If gaming performance REALLY matters...

CPU isn't that important.
In cases where you can get more frames from a faster CPU you're either

1. In hundreds of FPS  territory anyway.
2. have a VERY powerful videocard and running at relatively low settings.

Take just about any game with a 3090 and run it at 4K and it's effectively a 20-way tie for first place on the CPU.
And a 3060 at 1080p has a similar outcome.

[leadeater]

18 minutes ago, cmndr said:

If gaming performance REALLY matters...

CPU isn't that important.
In cases where you can get more frames from a faster CPU you're either

1. In hundreds of FPS  territory anyway.
2. have a VERY powerful videocard and running at relatively low settings.

The CPU is still important to a point, can also depend on the game too. Personally I play a lot of RTS and TBS games, CPU actually does matter here, some other games have strong AI (game AI not buzz word AI) CPU usage too. This will only trend up, same with graphical demands too of course.

 

Also I pay more attention to 1% and 0.1% lows than averages, personally I'd rather lose out a bit in average for a smaller distribution of frame times, but that's just me.

 

Also 4K remains in my opinion a silly resolution that doesn't give any benefit, gives the 3090's of the world a purpose though heh.

 

But this is why it wont matter, performance will be the "same" even if CPU bound Zen 4-3D vs Raptor Lake so who cares. Hopefully there will be more than a single product next time around with 3D V-Cache so buying a 12600k/12700k for less than a 5800X3D giving the same performance is not the obvious answer  

[KaitouX]

6 hours ago, cmndr said:

Zen 3 currently ties or slightly edges out ADL at MT while using substantially less power in most cases.

It could very well be the case that Zen 3 wouldn't beat RaptorLake though.

It's very probable that Zen 4 (at least without 3dVcache) won't have the ST crown vs RaptorLake but WOULD have a modest MT crown.

That really depends on the price point you're looking though. If you compare the 12900 to the 5950X, the 5950X is often better in MT performance, but both MSRP and prices for 3 months after Alder Lake was released put the 12900 in a huge price advantage, while current prices are close if you consider the 12900F, when you look at the 5900X and 12700 even now it greatly favors Intel, with the 12700F being almost $100 cheaper, while it was close to $200 for a few months. The 12600K and 5600X/5800X in particular were a ridiculously bad match up for AMD, with the 12600KF being available for $260, much cheaper than the 5800X for similar performance, and much faster than the 5600X for similar prices(before the price cuts), currently it's cheaper than the 5700X for better performance.

Without big price/core count changes, AMD has no chance of competing in value, flagship MT performance they are going to be competitive without a doubt, the issue is with the price and lower SKUs. Basically if prices doesn't change the 13900K price competitor isn't the 7950X, but rather the 7900X and so on.

[cmndr]

6 minutes ago, KaitouX said:

That really depends on the price point you're looking though. If you compare the 12900 to the 5950X, the 5950X is often better in MT performance, but both MSRP and prices for 3 months after Alder Lake was released put the 12900 in a huge price advantage, while current prices are close if you consider the 12900F, when you look at the 5900X and 12700 even now it greatly favors Intel, with the 12700F being almost $100 cheaper, while it was close to $200 for a few months. The 12600K and 5600X/5800X in particular were a ridiculously bad match up for AMD, with the 12600KF being available for $260, much cheaper than the 5800X for similar performance, and much faster than the 5600X for similar prices(before the price cuts), currently it's cheaper than the 5700X for better performance.

Without big price/core count changes, AMD has no chance of competing in value, flagship MT performance they are going to be competitive without a doubt, the issue is with the price and lower SKUs. Basically if prices doesn't change the 13900K price competitor isn't the 7950X, but rather the 7900X and so on.

Both AMD and intel have healthy per-unit margins right now. This means that price is pretty flexible. Both companies will aim to price to optimize a good balance of per-unit profit and market share.

There's no reason to think that there will be a huge gap in prices relative to performance unless there's some sort of VERY interesting "specialization" or edge case that favors a high value customer segment.

 

Also part of the reason why ADL is cheaper is that its complementary goods (motherboards) are more expensive. In practice consumers buy platforms.
I suspect that AM5 motherboards will cost a bit more than AM4 boards, so you can expect prices to level out somewhat.