Leaked Patch Confirms AMD Zen Will Have 32 Cores Per Socket?

[Zarathustra[H]]

And that is how it used to be see how well Battlefield 4 runs on a PS4 it can only get better. And you still want to keep up the serial aspect of it then BF4 simply would not have existed on consoles at all ..

What you don't seem to understand is that not all code lends itself to multithreading. It's elementary computer science, not just a desire to not do it.

There is always a percentage of code that simply does not lend itself well to threading, that's why the Amdahl's law chart has lines for different percentages.

Most game code tends to have a very high percentage of code that can not be threaded due to all the conditional outcomes and risks of thread locking where two (or more) threads becone stuck waiting g for each other. They often find ways around this by instead of truly multithreading the game code, splitting it up in multiple instances of serial code (like, sound engine in one thread, GPU supporting process in another, game logic in a third, etc.)

This helps you take advantage multiple cores better than just running one serial thread, but you are still running multiple serial threads that are either non-trivial or completely.impossible to split up further than they already are, and once it is done might actually reduce performance rather than improve it.

Because of this, for most titles, threading has pretty much gotten as good as it ever will at this point.

Tou are correct. BF4 is a great example of threading what CAN be threaded (like the physics engine and stuff like that) and splitting everything else up into smaller serialized chunks so that it distributes the load over multiple cores.

Some titles also lend themselves better to multithreading in the game engine, Like Civ5, as there are many semi-independent calculations between rounds, but other titles can't use this approach.

Unless the underlying theories of how software logic works change somehow, I don't see how we are ever going to improve threading of game titles beyond the level of a well threaded current title.

 

[Pieter3dnow]

Are you trying to say that I don't know what I'm saying because software that is serial can not function on more cores. While the trade off would be that things that can run on more cores should suffer from it, gaming can benefit greatly from this why should we be stuck we been stuck for a long time already now.



No one there ever said anything about problems that better parallel code would be a feature that makes no sense because of the things you said. And they all are doing rather complex games ....

 

[jwcalla]

Are you trying to say that I don't know what I'm saying because software that is serial can not function on more cores. While the trade off would be that things that can run on more cores should suffer from it, gaming can benefit greatly from this why should we be stuck we been stuck for a long time already now.

No one there ever said anything about problems that better parallel code would be a feature that makes no sense because of the things you said. And they all are doing rather complex games ....

He explained pretty well what the issues are. A game is, unfortunately, a very synchronized piece of software. And that synchronization is expensive. But there are some areas of improvement that the new APIs offer. Just don't expect miracles going forward.

 

[Zarathustra[H]]

Are you trying to say that I don't know what I'm saying because software that is serial can not function on more cores. While the trade off would be that things that can run on more cores should suffer from it, gaming can benefit greatly from this why should we be stuck we been stuck for a long time already now.



No one there ever said anything about problems that better parallel code would be a feature that makes no sense because of the things you said. And they all are doing rather complex games ....

Whatever man.

All I'm saying is, don't expect an FPS to be better threaded than BF4 anytime soon...

...or ever.

 

[Zarathustra[H]]

He explained pretty well what the issues are. A game is, unfortunately, a very synchronized piece of software. And that synchronization is expensive. But there are some areas of improvement that the new APIs offer. Just don't expect miracles going forward.

Exactly. Improves API's can help better distribute the portion of the CPU load that resutls from the GPU/API directly over more cores, but unfortunately that is a very small portion of the overall CPU load in a typical gane., so it won't make a huge difference either way.

 

[JustReason]

Exactly. Improves API's can help better distribute the portion of the CPU load that resutls from the GPU/API directly over more cores, but unfortunately that is a very small portion of the overall CPU load in a typical gane., so it won't make a huge difference either way.

Also keep in mind a great deal of what he is seeing and what makes these discussions difficult, is that with DX12 alleviating a few of the bottlenecks associated with DX11 CPU to GPU communication we now have the opportunity to use some of these before dormant resources which, not necessarily adding FPS (frames per second not first person shooter), would add more immersion to games. Hence why I don't think we will see FPS go from 60 to 120 but rather 60 to 75 with quite abit more of some things per frame.

 

[Pieter3dnow]

To be precise I'm never talking about a frames per second boost, I'm talking about the overall status of gaming and their engines , DX12 proves that there is so much to be won there like what Mantle did If you watch the APU13 video presentation (Oxide) you can see that not only due to overhead there is a lot of free cpu time now that close to metal api are here.

AMD Zen Rumours Point to Earlier Than Expected Release

If you check Phenom II x4 performance and see what it does for 8 core performance. Now that the doors have been opened I can't imagine for the API to reach a cap already after 8 cores. Again i have to stress that in the APU13 dev Q&A no one was saying anything about a hard cap for certain amount of cores beside the comment about DX11 stopping after 5 cores and 6 cores sometimes giving worse performance surely after that comment being made they would say something relevant about Mantle having a slightly higher cap....

 

[Zarathustra[H]]

Also keep in mind a great deal of what he is seeing and what makes these discussions difficult, is that with DX12 alleviating a few of the bottlenecks associated with DX11 CPU to GPU communication we now have the opportunity to use some of these before dormant resources which, not necessarily adding FPS (frames per second not first person shooter), would add more immersion to games. Hence why I don't think we will see FPS go from 60 to 120 but rather 60 to 75 with quite abit more of some things per frame.

I feel like this is a highly optimistic prediction. The optimizing of CPU load resulting from CPU to GPU type load is a very small deal. I don't even think the change will be noticable on a high end rig. You are probably going to have a less than 5% impact on CPU load that currenlty isn't the bottleneck anyway.

Where it becomes more significant is on the very low end power saving mobile devices, where every little bit helps and is crucial.

 

[Mackintire]

Last leak I saw it was a two 8 core processors glued together on the same die... like Pentium D. With SMT (Hyperthreading) it shows up in windows as 32 cores.

 

[Mackintire]

As I understand it AMD worked on a 8 core and a 4 core processor and then binned the resulting variants to fit all possible models.

From a KISS perspective this might be a good thing.

 

[Deathroned]

I really hope AMD isn't trying to launch a 32 core processor for the consumer market.. general task/gaming on an intel xeon 14C/28T is pretty abysmal due to lack of proper software. Virtualization on the other hand..

Though on the Opteron side, I can barely give away the darn things.

Do you even edit your cfg files bro. Joking aside you should check our venturi here on hardforum he games on a dual xeon[e5 2699v3] and quad titan x

 

[tangoseal]

I would love more cores. SInce the future of software is parallel threads might as well max them babies out.

 

[KazeoHin]

Do you even edit your cfg files bro. Joking aside you should check our venturi here on hardforum he games on a dual xeon[e5 2699v3] and quad titan x

I can say first hand that my 14c28t Xeon is just fine for how I use it. Multithreaded content creation applications eat up as many threads as you can give them. Games on the other hand tend to be CPU bottlenecked due to low single-core power. I am still able to play MOST games at high-ultra settings at 4K. MGS5, Ryse, Dirt Rally... all maxxed to the top. When DX12 becomes more optimised and popular, I'll expect at least 8 of my cores to be utilised, which means streaming and multitasking while gaming is still definitely possible.

 

[tangoseal]

I can say first hand that my 14c28t Xeon is just fine for how I use it. Multithreaded content creation applications eat up as many threads as you can give them. Games on the other hand tend to be CPU bottlenecked due to low single-core power. I am still able to play MOST games at high-ultra settings at 4K. MGS5, Ryse, Dirt Rally... all maxxed to the top. When DX12 becomes more optimised and popular, I'll expect at least 8 of my cores to be utilised, which means streaming and multitasking while gaming is still definitely possible.

Well if DX12 really shines then my 3930K might have a rebirth as another 3 year lifespan extension. Its already wicked fast and benches higher than an I7-6700K as is. I am watching ZEN super close though and I will jump on it the second it comes out if it is really that big of a performance leap over my 3930K.

 

[cageymaru]

As I understand it AMD worked on a 8 core and a 4 core processor and then binned the resulting variants to fit all possible models.

From a KISS perspective this might be a good thing.

Initial offerings will be 8 & 6 cores it seems. Maybe HT on / off to differentiate the processors?
AMD Will Only Be Fabricating Octa-Core Dies For Zen x86 Processors - No Dual Core or Quad Core Flavors Initially

Original article.
http://www.bitsandchips.it/52-english-news/6955-no-dual-and-four-cores-zen-cpus-at-least-initially

 

[Pieter3dnow]

Initial offerings will be 8 & 6 cores it seems. Maybe HT on / off to differentiate the processors?
AMD Will Only Be Fabricating Octa-Core Dies For Zen x86 Processors - No Dual Core or Quad Core Flavors Initially

Original article.
http://www.bitsandchips.it/52-english-news/6955-no-dual-and-four-cores-zen-cpus-at-least-initially

Yeah I saw the link earlier on another forum. Only thing now is better performing gaming engines and were good to go

Would hope that AMD leaves those lower cores to the APU parts if they have 4 cores to start with you can always sell the ones where the gpu has problems ..

 

[a3venom]

If this is true and these are going to be priced around the i5,i7 series with as-expected IPC - i think they are in for a winner.

 

[cageymaru]

If this is true and these are going to be priced around the i5,i7 series with as-expected IPC - i think they are in for a winner.

No. It is rumored that Intel is making a cheaper lineup just to compete against these.

 

[a3venom]

No. It is rumored that Intel is making a cheaper lineup just to compete against these.

AMD's quad cores will probably be little over i3s, and might perform like a 3570k. I dont think intel will have anything to beat that in that price range as that would just shake things up quite a bit. Quad cores will still be the mainstream for some time

 

[cageymaru]

AMD's quad cores will probably be little over i3s, and might perform like a 3570k. I dont think intel will have anything to beat that in that price range as that would just shake things up quite a bit. Quad cores will still be the mainstream for some time

AMD quad cores are going to be... APUs. Note how well they sell now. Of course they are getting updated; I just don't see them becoming "awesome".

 

[Pieter3dnow]

AMD's quad cores will probably be little over i3s, and might perform like a 3570k. I dont think intel will have anything to beat that in that price range as that would just shake things up quite a bit. Quad cores will still be the mainstream for some time

Can you show us numbers where you base this on? 4 core Zen APU is the same as a I3 ?

 

[a3venom]

Can you show us numbers where you base this on? 4 core Zen APU is the same as a I3 ?

I am just speculating --- let's say post Zen we have the following tier of CPUs where AMD will make super good sense. 4 core APU is not an attractive concept whereas a cheaper 4 core ivybridge-IPC CPU is very attractive.

If there are following price tiers of CPUs
i3 ~120
4 Core Zen ~ 170
i5 (mid range like 6500?) ~ 200
4 Core/8t Zen ~ 250
i5k Series (slightly above the 4c/8t core or same price) ~250
6 core/12t Zen = i7 4c/8t ~ 350
Intel 6 core / 12t (probably 150$ more)

In this case AMD is winning heavily on price and also on performance. Even if the AMD prices are higher than what i am speculating - that would be what everyone really wants.

 

[drescherjm]

I would be happy with those prices if the 40% IPC improvement is true and the clocks are at least similar to Intel for the same # of cores. I hope that they will still have ECC enabled. I would be looking at the 4C / 8t to replace my aging core2quad linux server / pvr. Although my windows workstation I will be considering at the server platform in 2017 since I have a 6 core / 12 threaded CPU currently.

 

[Mackintire]

More realistic would be:

2 core Zen ~80-130
i3 2 core~120
4 Core Zen low end ~145+
i5 (mid range like 6500?) ~ 200
4 Core/8t Zen ~ 145-280 depending on bin and clk
i5k Series (slightly above the 4c/8t core or same price) ~250
6 core/12t Zen ~ 300-350 depending on Bin
Intel 6 core / 12t (~380)
Intel 8 core /16t (~650)
Intel 10 core/20 t (~1100)

 

[Mackintire]

I would be happy with those prices if the 40% IPC improvement is true and the clocks are at least similar to Intel for the same # of cores. I hope that they will still have ECC enabled. I would be looking at the 4C / 8t to replace my aging core2quad linux server / pvr. Although my windows workstation I will be considering at the server platform in 2017 since I have a 6 core / 12 threaded CPU currently.

Zen apparently benches very close to Haswell in performance but the final clock speed is what is going to determine the final outcome. If the top Zen Quad core is going to be 3Ghz base w/turbo being 3.5Ghz , that puts it 20% behind the i5-6600 on 4 thread workloads and maybe on equal or slightly better ground on some higher thread applications.

Top speed of an earlier engineering sample was 3.2Ghz, so ~3.5Ghz release silicon is not unreasonable.

If the 6 core Zen manages a 2.7Ghz base clk with the same 3.5Ghz turbo it will still get beaten by the Intel 6 core i7-5820k by 20% as well.

The same 6 core Zen chip w/ 2.7clk base and 3.5Ghz turbo will be a tough sell against the i7-6700, because in most use cases the Intel chip will outperform the (example) Zen by 50% or more. Under 6+ thread load Zen might make a comeback and close the performance margins to less than 20%

The other side of the equation is how good will the APU/graphics portion be. It might be x2 higher than Intel's solution.

 

[Zarathustra[H]]

Zen apparently benches very close to Haswell in performance but the final clock speed is what is going to determine the final outcome. If the top Zen Quad core is going to be 3Ghz base w/turbo being 3.5Ghz , that puts it 20% behind the i5-6600 on 4 thread workloads and maybe on equal or slightly better ground on some higher thread applications.

Top speed of an earlier engineering sample was 3.2Ghz, so ~3.5Ghz release silicon is not unreasonable.

If the 6 core Zen manages a 2.7Ghz base clk with the same 3.5Ghz turbo it will still get beaten by the Intel 6 core i7-5820k by 20% as well.

The same 6 core Zen chip w/ 2.7clk base and 3.5Ghz turbo will be a tough sell against the i7-6700, because in most use cases the Intel chip will outperform the (example) Zen by 50% or more. Under 6+ thread load Zen might make a comeback and close the performance margins to less than 20%

The other side of the equation is how good will the APU/graphics portion be. It might be x2 higher than Intel's solution.

Yeah, what will be interesting to me is if the non-GPU parts will clock higher.


If you recall with APU's (I forget in which generation) they moved to a production process that was more GPU friendly, but also sacrificed max CPU clock. (can't recall details). I wonder if they will be able to scale clocks on the CPU portion of CPU only chips up higher by using a CPU dedicated process, or if they will be producing all of them, APU's and CPU's on the same process.

I am hopeful that the first one is true, just based on the fact that the FX CPU's are supposedly launching much sooner than the APU's. This means that they COULD be on separate processes.

For reference, here is my little linear interpolation based forward looking performance analysis from another thread (in which the OP was asking about video rendering performance)

I don't know what your video rendering looks like, but let's assume for a moment that it performs similarly to Cinebench 11.5:

A 6700k at max turbo of 4.2ghz scores 2.07 right now in the single core test. That's ~0.49/ghz.

A 7870k at max turbo of 4.1 scores 1.06 right now in the single core test. That's ~0.26/ghz.

So lets add the much talked about 40% to that figure. We get a predicted Zen performance of ~0.36/ghz. This is still only ~73.5% of the score of a Skylake core per ghz. Maybe AMD will be able to ramp up the clocks a little bit? Their current APU's are held back in clock because they have to use a GPU friendly manufacturing process. On pure FX chips this won't be an issue and maybe they'll be able to squeeze more clock speed out of them? If we assume this, and say maybe they can get a 4.8ghz turbo out of an FX based Zen, where does that leave us? Well, at 4.8Ghz, the predicted Zen score would be 1.74. This is a huge improvement, but it is still only 84% as fast as a current skylake.

That being said, a 1.74 score, ties Zen with a Haswell i7-4770K which is a respectable leap forward.

In other words, these calculations demonstrate what we have been guessing all along (or at least were, before the hype machine took over), that Zen will allow AMD to catch up to where Intel was a generation or two ago core for core. Being competitive with Haswell is actually better than I was expecting. I was thinking more along the lines of Ivy Bridge.

With your type of workloads (video rendering) you should scale very well with core count too, and if history is any judge AMD will release many core chips. Zen might just be the platform of choice for video rendering / encoding type workloads if they add many cores. (though I wouldn't count on consumer 32 core chips as have been rumored)

Now, this analysis is to be taken with TRUCKLOAD of salt as it is based on many assumptions and both educated and unedcuated guesses. Also, there is a lot of criticism of Cinebench as a benchmark, as it is believed to strongly favor Intel chips and not be a fair comparison. I used it because it is one of the few benchmarks that I can find results from that isolate single core results (as throwing in extra cores muddies the waters), and because you spoke of doing video rendering. The intent - however - is to put things in perspective, of what we can expect from Zen IF they live up to the official marketing slides from AMD.

 

[drescherjm]

Yeah, what will be interesting to me is if the non-GPU parts will clock higher.


If you recall with APU's (I forget in which generation) they moved to a production process that was more GPU friendly, but also sacrificed max CPU clock. (can't recall details). I wonder if they will be able to scale clocks on the CPU portion of CPU only chips up higher by using a CPU dedicated process, or if they will be producing all of them, APU's and CPU's on the same process.

I think this was more a combination of the node shrink (Intel has had the same issue since 22nm) and no longer using SIO.

John

 

[livfree]

Or are they going to have 16 'modules' and 32 cores?

 

[Pieter3dnow]

Or are they going to have 16 'modules' and 32 cores?

No modules this time around and Simultaneous Multi threading. 8 cores 16 threads

 

[livfree]

Damn, maybe AMD will have a chance to be worth while again.

 

[Mackintire]

On the server side they mentioned placing two units on one die just like the pentium D which would make the top configuration a 6+6 +HT or 4+4+HT.

Personally I think Zen is more about getting the product right and ripping the floor out from Intel on the low end through mainstream....and Zen + is where they will optimize to scale the # of cores and clk speed in hopes of taking back the 2P server market and continuing to grow market share.

 

[mwnn]

I hope they keep it simple personally.

I really don't think we need a bunch of redundant lower clocked or otherwise obscure models to pick from. 2-3 from each category would do me. Do we even need any dual core models if they can make an efficient quad?

Browsing through used Intel parts and hoping for a bargain, I'm already sick of looking through all the combinations of chipsets, sockets & numbering for very similar parts, without adding in the Xeons to complicate things further.

Gaming mostly with the odd bit of video conversion, archiving etc.

I'd like to see:

• Universal AM4 socket regardless of CPU cores (I hope they do make a desktop 32 core for the zealous/as a demo)

• Overclocking as standard as with the FX line

• 4 cores / 8 threads @ > 3GHz with Haswell-like IPC -> cheap quad cores for the masses

• 8 cores / 16 threads @ > 3Ghz with Haswell-like IPC -> as a better longer term option banking on it lasting 5+ years

• Efficiency in line or at least similar to Intel

• Undercut the current Intel pricing

Instant seller for me and hopefully more profitable for them than prior designs. And the stock cooler is no longer useless at last - that thing sounded like a jet engine on my Phenom II.

Do that and they've wiped out the overpriced used Intel i5 / i7 market to boot.

I honestly don't see how they could cock this up unless they get complete imbeciles to set the prices.

I've hung onto my Phenom II for so long that's it's going to be almost impossible for them to disappoint me, any decent chip should be a 2-3x jump.

 

[drescherjm]

• 4 cores / 8 threads @ > 3GHz with Haswell-like IPC -> cheap quad cores for the masses

I am not sure what you consider cheap, however I do not expect these to be less than $175. The 4 core / 8 threaded CPUs will compete against i5s.

 

[a3venom]

I am not sure what you consider cheap, however I do not expect these to be less than $175. The 4 core / 8 threaded CPUs will compete against i5s.

4c/8t Haswell-ish IPC Zen should rape an i5

 

[drescherjm]

4c/8t Haswell-ish IPC Zen should rape an i5

Why/ How? An i7 does not and the 4 core / 8 threaded Zen will in no way be a Skylake / Kabylake i7.

 

[KazeoHin]

Personally, my preferred SKU lineup would be:

Zen Casual:
4 core +/- HT, 20-45w $60-$90 (performance similar to a modern i3)

Zen Mainstream:
8 core +/- HT, 45w $120-$200 (performance similar to a high-end i5)

Zen Enthusiast:
12 core + HT, 65-95w $250-$350 (performance surpassing a 6-core i7)

Zen Extreme:
16 core + HT, 95-120w $400-$600 (performance close to a 10-core i7)

Going above and beyond and offering unlocked Server-spec AM4 chips with >16 cores would get an INSTANT sale from me.

Something like

Zen Pro:
24 core + HT, 95w $600-$750 (performance unlike any unlocked chip available)

Zen Pro Extreme:
32 Core + HT, 145w $1200 (performance unlike any unlocked chip available)

 

[Pieter3dnow]

Little bit baffled by the wattage you link 8 cores 45 Watt ? 2ghz per core then ? Maybe use a find and replace HT with SMT since HT sounds similar but really is not the same thing.

For AM4 just 8 cores with SMT. If there would be a 16 core part (or some other number higher then 8) for AM4 depending on how the situation is with games I would not mind buying it even if it had a premium price tag.

 

[mwnn]

I am not sure what you consider cheap, however I do not expect these to be less than $175. The 4 core / 8 threaded CPUs will compete against i5s.

That's £119.

I paid that for my Phenom II a long time ago.

The FX 8350 is going for £130 currently on Amazon UK.

With a 6600K being a whopping £199 (even used Intel is going for £100+ eBay)


I'd be very happy with that estimated price - quite reasonable to be honest - even happier if it gave an i5 a good thrashing.

Zen 4 Core / 8 threads facing off against i5 6500, previous gen models and below

Zen 8 core / 16 threaded models facing off against i5 6600k and above

But It's all theoretical at this point isn't it =D


I'd consider buying anything less as a bit of waste - especially in cases where the price of a board exceeds the price of the CPU!

I think chips like the Athlon X4 860 are completely useless and I'm not sure why they bother.
AMD Athlon X4 860K Quad-Core 3.7GHz 4MB Socket FM2+ 95W Desktop Processor: Amazon.co.uk: Computers & Accessories

Bargain basement machines get an APU.

Personally, my preferred SKU lineup would be:

Zen Casual:
4 core +/- HT, 60w $60-$90 (performance similar to a modern i3)

Zen Mainstream:
8 core +/- HT, 95w $120-$200 (performance similar to a high-end i5)

Zen Enthusiast:
12 core + HT, 125w $250-$350 (performance surpassing a 6-core i7)

I'd be happy with these numbers although maybe the TDP numbers are a bit optimistic knowing AMD? So I've changed them =P

 

[Araxie]

Little bit baffled by the wattage you link 8 cores 45 Watt ? 2ghz per core then ? Maybe use a find and replace HT with SMT since HT sounds similar but really is not the same thing.

they are the same.. just that intel name it Hyper-threading but they work exactly the same at thread level of execution, so 2 logic processors per core sharing individual core resources, the implemented "SMT" will work the same as intel Hyper-threading, you are probably referring to Temporal Preemptive Multitasking which just operate one instruction per thread instead of multiple page tables, tasks and I/O as hyper-threading do and AMD SMT will also do.. at the end of the day, zen will work exactly as any other intel hyper-threading implementation since Ivy Bridge up to broadwell but I don't think it will have the Skylake Hyper-threading optimizations which have improved front-end, more executions units, out of order and wider retirement with ADX instructions..

Also I hope your not confusing SMT with CMP..

so at the end they will be the same 8 physic + 8 logic. and not 16 fully true cores.

 

[Pieter3dnow]

they are the same.. just that intel name it Hyper-threading but they work exactly the same at thread level of execution, so 2 logic processors per core sharing individual core resources, the implemented "SMT" will work the same as intel Hyper-threading, you are probably referring to Temporal Preemptive Multitasking which just operate one instruction per thread instead of multiple page tables, tasks and I/O as hyper-threading do and AMD SMT will also do.. at the end of the day, zen will work exactly as any other intel hyper-threading implementation since Ivy Bridge up to broadwell but I don't think it will have the Skylake Hyper-threading optimizations which have improved front-end, more executions units, out of order and wider retirement with ADX instructions..

Also I hope your not confusing SMT with CMP..

so at the end they will be the same 8 physic + 8 logic. and not 16 fully true cores.

Nope I am sure that SMT is a different implementation then what Intel calls Hyper Threading. What you are saying that there enough commonalities to blur them. I am sure that you can find yourself a few examples where they discuss the difference between SMT and HT that they approximately do the same job is something we both know.
Usually processors are designated as cores and threads