Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[adroc_thurston]

AVX-VNNI support

They already support both AVX512-VNNI and AVX512-BF16.

 

[Timorous]

100% performance increase per core going from Zen3 -> Zen5, really? What do you base that on?

Also, Zen5 may very well have memory controller and other CPU design improvements + more cache, making it less memory bandwidth hungry than Zen3, all else equal.


Of course, more memory bandwidth is always better. But in several workloads the difference was not that big. And note that we’re talking about twice the amount of cores (64), and half the bandwidth (DDR4 3200). So with 32C and DDR5 6400 you’ll have 4x the memory bandwidth per core compared to that, when using the same number of channels.

In product comparing mid range ThreadRipper to top range desktop I don't see why not, TR tends to run at lower all core clocks and even if Zen 5 has a small clock regression from Zen 4 it will still have a large clock speed uplift vs Zen 3 and of course AVX 512 will play a part. Some workloads will still be faster on the 32c Zen3, blender, CineBench etc probably but there will be plenty of other workloads that will be far far faster on the Zen 5.

The 7950X already matches the 5965WX and is nipping on the heels of a 3990X when you take a plethora of workloads and do a geomean.

He also tested memory scaling on the 5965WX and found that in some workloads there is an astronomical difference in performance. In others not so much but if the 5965WX was only using 4 channels it would be closer to a 14900K in this chart on the basis of that scaling test.

So do I think a 16c Zen 5 desktop part could match a 5975WX in this test suite, yes I think it could give or take a bit. The 5975WX is at best 33% faster than the 5965WX and if you factor in the slightly lower all core clocks, the lower single core clocks and the tests that don't scale with more cores you are probably looking at 20-25% faster than the 5965WX as tested. I think a 9950X could manage that over the 7950X in this test suite.

I also don't think a 32c Zen 5 desktop part would give you much. it would need Zen5c which means you are going to take a huge hit in ST and lightly threaded performance and you are going to be bandwidth constrained in lots of other workloads meaning the only chart it might be faster in is Blender / CineBench and if that is all you care about why not GPU render which is even faster still? Same issue with a 24c hybrid of Zen5 and Zen5c although you don't have the ST drawbacks. You would still lose out in moderately threaded stuff to gain a little bit in Blender / Cinebench.

 

[Timorous]

If (according to sources like OC3D, GN, TS) a 24c CPU from another brand matches a 7950x in MT while having a 1% difference between DDR4 and DDR5, then there is no basis to say memory bandwidth is an issue in the first place. A 24c Zen 4 CPU would be fine on DDR4 and a desktop Zen 5 CPU with 32c let alone 24c would also still be fine on DDR5 for the market segment that its targeting.

Those sources are not great for actual productivity testing. Puget is pretty good and phoronix is very broad so covers pretty much all bases.

The only segment that might benefit from a 24c zen5 / zen5c setup are people with a mainly ST workflow with a heavy MT step at the end that are not doing it for money, or are not doing it for much money.

 

[StefanR5R]

Here's an article. 1, 2, 4, 8, memory channels and 32, 64 cores tested.

AMD Threadripper PRO Memory Channel Performance Scaling
https://www.pugetsystems.com/labs/articles/amd-threadripper-pro-memory-channel-performance-scaling/

Thanks for finding this. Though in several of these tests, 5965WX (24 cores) is as fast or faster than 5995WX (64 cores). Maybe a 16-core CPU would be even faster...

(And as far as image rendering tests are concerned: Yes indeed, rendering small scenes does not require memory bandwidth. But what if somebody wants to render large scenes, or a large number of small scenes? The answer is of course that they do it on GPUs, not on high core count CPUs.)

Edit,

[Phoronix] also tested memory scaling on the 5965WX and found that in some workloads there is an astronomical difference in performance. In others not so much but if the 5965WX was only using 4 channels it would be closer to a 14900K in this chart on the basis of that scaling test.

Thanks for this find!

 

[Timmah!]

I also do not see a reason to offer more. If you do need more mt than what 16 Zen5 cores can offer then chances are you are being paid for it and you should probably run the numbers on a TR system because long term it will pay for itself.

Consider the chances you are NOT being paid for it, which IMO is true for much bigger part of the world than it isnt. In US, cashiers at stores can be paid as much money per hour, as some people with master degrees in some parts of Europe.

 

[bender250]

Similar bandwidth arguments were around before release of 16 core Zen 2. Both the Puget and Phoronix reviews illustrate that there are workloads where memory bandwidth does not play a significant role (Phoronix - compilation, Puget - rendering). So there are some of us running such workloads and waiting for AMD to release CPUs with more cores, as we know we would be benefiting from it. So the argument of "more cores with 2CH memory are not needed" is not valid *in general* - please do not judge for all of us. The main reason that AMD is not increasing core count is lack of competition and their own profits cannibalization, but it is not that there would not be users of such systems.

For me would 32 core Zen 5 be worth easily $2k, maybe 2.5k. Only alternatives I have are second-hand Epyc CPUs or maybe Siena. Those are the markets such hypothetical 32 core would cannibalize.

 

[SteinFG]

For me would 32 core Zen 5 be worth easily $2k, maybe 2.5k. Only alternatives I have are second-hand Epyc CPUs or maybe Siena.

What's wrong with threadripper 7970X? (or 9970X for Zen 5 when it comes out)

 

[Timorous]

Similar bandwidth arguments were around before release of 16 core Zen 2. Both the Puget and Phoronix reviews illustrate that there are workloads where memory bandwidth does not play a significant role (Phoronix - compilation, Puget - rendering). So there are some of us running such workloads and waiting for AMD to release CPUs with more cores, as we know we would be benefiting from it. So the argument of "more cores with 2CH memory are not needed" is not valid *in general* - please do not judge for all of us. The main reason that AMD is not increasing core count is lack of competition and their own profits cannibalization, but it is not that there would not be users of such systems.

For me would 32 core Zen 5 be worth easily $2k, maybe 2.5k. Only alternatives I have are second-hand Epyc CPUs or maybe Siena. Those are the markets such hypothetical 32 core would cannibalize.

There are users for everything. TR is already super niche in DIY and even as a pro workstation lineup and I think higher core count Ryzen desktop parts would also be super niche.

It all comes down to the fact the juice is not worth the squeeze.

 

[bender250]

What's wrong with threadripper 7970X? (or 9970X for Zen 5 when it comes out)

Total costs: the $2-2.5k for 32 core Zen 5 translate to $1.5k-2k in 32 core Zen 4 and the motheboards cost also additional $500. I rather built two machines with 7950X for $3k.

There are users for everything. TR is already super niche in DIY and even as a pro workstation lineup and I think higher core count Ryzen desktop parts would also be super niche.

It all comes down to the fact the juice is not worth the squeeze.

That is surely part of the consideration that AMD is doing. I still expect the higher-products cannibalization being more relevant to AMD's economical decision.

But same logic is valid to 16 core from Zen 2 lineup. The reason why they released it back then was that they needed the complete performance lead, while now such a 32 core part is not needed. I am staying curious if Intel will be the one to push the core count, whether the leaks about second batch of Arrow Lake with 8+32 cores are sound, as I expect that would push AMD into releasing higher core parts.

To the main topic of the thread, if people are betting the performance lead, here is my guess (based on just reading this thread). 18% single-threaded total performance improvement over Zen 4, and 15% in multi-threaded load: higher gain in FP (wider core), but lower gain in INT loads (consumption) - but I wish hyping by @adroc_thurston of the front-end is true and I am wrong. Zen 4 already increased MT performance a lot, so I think it will be more challenging without significant manufacturing node jump. Where is the guess from @inf64? What if reviews are out in 3 months, you do not wish to miss the chance extend your record of successful guesses .

 

[inf64]

@bender250

I did post a few times but I didn't have time to do a bigger post like in the past . I will try to do it before this thing launches, but this will be the toughest one to predict and probably where my lucky streak ends

 

[adroc_thurston]

18% single-threaded total performance improvement over Zen 4, and 15% in multi-threaded load: higher gain in FP (wider core), but lower gain in INT loads (consumption)

See this is below of what Zen3 did and you can't deliver below of what Zen3 did with a fatter tock.

 

[Timorous]

See this is below of what Zen3 did and you can't deliver below of what Zen3 did with a fatter tock.

Sure you can. You don't design for it ofc but it is entirely possible to miss targets, look at Intel for the last 10 years.

 

[adroc_thurston]

Sure you can

Not at AMD (or anywhere else that matters) really.

but it is entirely possible to miss targets

Not to the degree that matters here.
It's a bigger tock than Zen3 made to do bigger things.

 

[Fjodor2001]

I am staying curious if Intel will be the one to push the core count, whether the leaks about second batch of Arrow Lake with 8+32 cores are sound, as I expect that would push AMD into releasing higher core parts.

^^^^ This. AMD is intentionally sandbagging the core count on desktop CPUs now. This time I think it'll be Intel that leads the dance to higher core count, and AMD will then reluctantly follow.

 

[adroc_thurston]

This time I think it'll be Intel that leads the dance to higher core count, and AMD will then reluctantly follow.

They don't have to do anything.
Atom spam doesn't solve any of the problems Intel has.

 

[gdansk]

And are they making a new IOd with more GMI links? Otherwise a 32 core part would have to be all dense cores and not worth buying in the slightest. An emergency edition product not worth making.

 

[StefanR5R]

But same logic is valid to 16 core from Zen 2 lineup. The reason why they released it back then was that they needed the complete performance lead, while now such a 32 core part is not needed.

Yet even though they maybe didn't need to, AMD did release 24-, 32-, and 64-core CPUs (and actually also a 96-core CPU) to the DIY market, in the Zen 4 generation. Just because the price-per-core of these CPUs is higher than, say, that of a run-of-the-mill 8-core CPU, and they only work in top-tier mainboards¹ doesn't mean that these products don't exist or cannot be bought. Yes, their price level is certainly somewhat influenced by what the direct competitor is offering.

¹) Actually, AMD went through hoops to create an extra stripped-down platform for these CPUs, with 2/3rds of the memory channels disabled, xGMI capability disabled, and several PCIe links disabled for the purpose. Yet people still complain about it, or keep ignoring its existence.

What AMD did not do however was to create a special IOD exclusively for a 24-core part that would fit into AM5. Duh.

Both the Puget and Phoronix reviews illustrate that there are workloads where memory bandwidth does not play a significant role (Phoronix - compilation, Puget - rendering).

I commented earlier on how CPU image renderer benchmarks fail to reflect real applications. Now, code compilation: If somebody works with a large code base on a desktop computer, i.e. a single-user PC, then they compile the code just incrementally much of the time. Which has low parallelism. IOW, high parallelism is rare for the use case of code compilation, making it a niche.

Or are you all like me and still use Gentoo Linux? (I do so on a 4c/8t Haswell though.)

 

[yottabit]

Total costs: the $2-2.5k for 32 core Zen 5 translate to $1.5k-2k in 32 core Zen 4 and the motheboards cost also additional $500. I rather built two machines with 7950X for $3k.


That is surely part of the consideration that AMD is doing. I still expect the higher-products cannibalization being more relevant to AMD's economical decision.

But same logic is valid to 16 core from Zen 2 lineup. The reason why they released it back then was that they needed the complete performance lead, while now such a 32 core part is not needed. I am staying curious if Intel will be the one to push the core count, whether the leaks about second batch of Arrow Lake with 8+32 cores are sound, as I expect that would push AMD into releasing higher core parts.

To the main topic of the thread, if people are betting the performance lead, here is my guess (based on just reading this thread). 18% single-threaded total performance improvement over Zen 4, and 15% in multi-threaded load: higher gain in FP (wider core), but lower gain in INT loads (consumption) - but I wish hyping by @adroc_thurston of the front-end is true and I am wrong. Zen 4 already increased MT performance a lot, so I think it will be more challenging without significant manufacturing node jump. Where is the guess from @inf64? What if reviews are out in 3 months, you do not wish to miss the chance extend your record of successful guesses .

Are we doing The Price is Right rules?

I’ll guess +35% ST and +22% in MT (power limited) with PBO bringing it up to +30% MT

 

[adroc_thurston]

And are they making a new IOd with more GMI links?

No.
cIOD is the same.

 

[Fjodor2001]

Otherwise a 32 core part would have to be all dense cores and not worth buying in the slightest. An emergency edition product not worth making.

8P + 2x16E has been rumored for second batch of Arrow Lake. Would be very good for a lot of the MT use cases being discussed here.

 

[jpiniero]

Would be very good for a lot of the MT use cases being discussed here.

But not the target market which is games.

 

[Fjodor2001]

But not the target market which is games.

You mean the subset of the market that you happen to be interested in. But not everyone only plays games.

 

[adroc_thurston]

But not everyone only plays games.

DIY CPU is pretty much games.

 

[TESKATLIPOKA]

But not the target market which is games.

For games a 16C Zen5 is also not the best option thanks to 2 CCDs.

 

[gdansk]

8P + 2x16E has been rumored for second batch of Arrow Lake. Would be very good for a lot of the MT use cases being discussed here.

Good for Intel but it doesn't suddenly make a 32 Zen 5C product desirable. If IOd is the same then 24 cores mismatched between Zen 5 and Zen 5C seems the maximum useful configuration that doesn't give up 1T performance.