Zen 2 rumors

Neapolitan6th

[H]ard|Gawd
Joined
Nov 18, 2016
Messages
1,182
I saw that too. Safe to say many would be very happy if this news were true.

An increase of 10-15% IPC would put Ryzen much closer to Skylake IPC, no?

Higher clocks from a 7nm process would be great too. 4.4ghz is achievable if you are lucky on their 12nm process.

Perhaps, 4.5, 4.6ghz would be a fairly realistic boost to expect from 7nm? Anything more would be icing on the cake.

As for an 8 core CCX, that is hard to say. I'd be more interested in a 6 core CCX with the other CCX being Vega integrated graphics.

That way AMD could push 6 core APUs into OEM mainstream desktops and potentially laptops. I imagine OEMs would recognize the marketing value of 6 cores over 4 cores.

To make the 6 core APU a successful move, I would suggest AMD replaces the current 2200G with a 6 core variant at the same price.
 

Gideon

2[H]4U
Joined
Apr 13, 2006
Messages
3,151
10 to 15% would put Zen above Skylake IPC. Tho the rest would be dependent on the clock speeds they can achieve. I am going to assume 4.8+ might be possible on 7nm. The CCX will likely be 6 cores but it is possible they can do 8 which would be very nice boost in performance. 6 core APU is likely why most think it will only be 6 cores per CCX and would be a good performance bump for laptops. If everything is true then Intel is in for a rough year ahead, but plenty of things really come down to how smooth the 7nm process goes.
 

rtangwai

[H]ard|Gawd
Joined
Jul 26, 2007
Messages
1,369
Does it really make sense to up the core count per CCX? Infinity Fabric was set up so that adding more CCX's would be simple, allowing AMD to scale pretty much at will (see Threadripper/EPYC).

I'm sure it makes sense to up the core count per CCX at some point, but is Zen2 really a good place to start? Why not leverage Infinity Fabric and concentrate on higher clockspeeds/IPC and lower power consumption?

If Zen2 turns out to be a true Intel-slayer ala Athlon then upping core counts would be logical for Zen2+. Upplng clockspeeds/IPC is hard enough, why make it harder by diverting resources?

Ironically enough I'm suggesting AMD follows Intel's tick-tock strategy.
 

Nobu

[H]F Junkie
Joined
Jun 7, 2007
Messages
8,843
Does it really make sense to up the core count per CCX? Infinity Fabric was set up so that adding more CCX's would be simple, allowing AMD to scale pretty much at will (see Threadripper/EPYC).

I'm sure it makes sense to up the core count per CCX at some point, but is Zen2 really a good place to start? Why not leverage Infinity Fabric and concentrate on higher clockspeeds/IPC and lower power consumption?

If Zen2 turns out to be a true Intel-slayer ala Athlon then upping core counts would be logical for Zen2+. Upplng clockspeeds/IPC is hard enough, why make it harder by diverting resources?

Ironically enough I'm suggesting AMD follows Intel's tick-tock strategy.
If you can do it without hurting yields or clockspeed substantially, then it makes plenty of sense. More cores per CCX means less jumping between CCXs and fewer cache misses = better performance in general.
 

KazeoHin

[H]F Junkie
Joined
Sep 7, 2011
Messages
8,721
I imagine a 6-core CCX being almost garenteed at this point, with all the rumours of Intel developing an 8-core 115X i7, I think AMD would be stupid to try to market against them with a Ryzen 8 core... A 12 core on the other hand...
 

Ebernanut

[H]ard|Gawd
Joined
Dec 15, 2010
Messages
1,686
At this point I'd be more interested in IPC and/or clock speed improvements but more cores would be nice and I think it's likely. If they increase the number of cores per CCX I wouldn't be surprised to see them offer more core count options and if they don't it might allow them to more aggressively bin the top CPU for each line, assuming the high end is fast enough for a slower version to make sense and that they don't just cannibalize them for TR.

The best part of these rumors is that I can enjoy hearing about them despite having just built a new system since they should drop right into my x470 board with a BIOS update.
 

Pieter3dnow

Supreme [H]ardness
Joined
Jul 29, 2009
Messages
6,784
I would rather see something as 12 cores rather then 16 with enough headroom for the range to stay within a certain TDP and still allow higher clocks.
If they found a better option for the CCX performance that would be welcome as well.
If not then another 8 core would not be that bad anyway.
 

tayunz

[H]ard DCoTM October 2018
Joined
Sep 27, 2004
Messages
1,343
I read that Rome (EPYC Zen 2) tops out at 48 cores, so doesn't this point to 6 core CCX?
 

tangoseal

[H]F Junkie
Joined
Dec 18, 2010
Messages
9,425
10 to 15% would put Zen above Skylake IPC. Tho the rest would be dependent on the clock speeds they can achieve. I am going to assume 4.8+ might be possible on 7nm. The CCX will likely be 6 cores but it is possible they can do 8 which would be very nice boost in performance. 6 core APU is likely why most think it will only be 6 cores per CCX and would be a good performance bump for laptops. If everything is true then Intel is in for a rough year ahead, but plenty of things really come down to how smooth the 7nm process goes.

This ...

It would put it above skylake. As it stands Zen + is less than %15 away from skylake IPC.

I am excited at the prospect of Zen 2 and their 16 core desktop parts. Not even HEDT and we will get 16 cores. This is fantastic rumor news.
 

Sulphademus

Limp Gawd
Joined
Mar 18, 2010
Messages
393
I'd also prefer more IPC and MHz over more cores right now but I don't do video editing or CAD rendering.
 

juanrga

2[H]4U
Joined
Feb 22, 2017
Messages
2,804
They are just rumors but I thought I would post a link to the article.

https://www.overclock3d.net/news/cp...a_10-15_ipc_boost_and_up_to_16_cores_on_am4/1

False rumor? They seem to obtain the 10--15% IPC from some ancient slide about performance

steamrollerportada2-1200x600.jpg


And the 8-core CCX rumor is also false. How in hell do you expect AMD to throw twice more cores and bigger cores in the same socket, whereas feeding all the cores with a dual channel?
 

cyberguyz

Gawd
Joined
Aug 28, 2014
Messages
710
And the 8-core CCX rumor is also false. How in hell do you expect AMD to throw twice more cores and bigger cores in the same socket, whereas feeding all the cores with a dual channel?

One memory channel per CCX. 2x 4-core CCX in Ryzen = 2 channel memory. 4x CCX in Threadripper = quad-channel memory
Core count per ccx does not dictate the number of channels it needs. Granted caching would need to be expanded to handle the higher core load, but that is well within the realm of possibility. I would think the higher level of multiplexing would cause some headaches but, well they have engineers to figure that one out.
 

Brackle

Old Timer
Joined
Jun 19, 2003
Messages
8,197
from what I remember IPC wasn't all that behind when clocked at the same speed. I can't remember where I saw the benchmarks, but AMD is behind in the IPC department no doubt. But IMO the issue is clock speed. Imagine if they could hit 4.8ghz on Ryzen+ and left IPC alone, that to me would be a huge increase in speeds compared to intel.

The reason Intel has the advantage in games (only talking about games) is mostly clock speed. Sure there is an IPC lead, but clock speed is the biggest factor.
 
Last edited:

schmide

Limp Gawd
Joined
Jul 22, 2008
Messages
366
I'm sure some people know this but won't say anything because of ???

If AMD improves their ability to retire data, they will receive a decent improvement. Additionally if they improve their FP rate which is behind intels, benefit.

The long and short of it is, it much easier for AMD to close the gap to intel, than it is for intel to surge further ahead.
 

kirbyrj

Fully [H]
Joined
Feb 1, 2005
Messages
30,216
I imagine a 6-core CCX being almost garenteed at this point, with all the rumours of Intel developing an 8-core 115X i7, I think AMD would be stupid to try to market against them with a Ryzen 8 core... A 12 core on the other hand...

One that runs on currently available boards with a bios update too. Fuck you, Intel after your Z270 fiasco.
 

schmide

Limp Gawd
Joined
Jul 22, 2008
Messages
366
Do you guys really think there will be more than 4 cores in a CCX?

A quick reminder of how the interconnect complexity grows.

The number of pathways between n objects is the combination C(n,2).

So for objects

2 = 1
3 = 3
4 = 6
5 = 10
6 = 15
7 = 21
8 = 28
16 = 120

One thing I've noticed about AMD is they don't give up on a design easily. Anything other than 4 core CCX would basically be a new chip.
 

tayunz

[H]ard DCoTM October 2018
Joined
Sep 27, 2004
Messages
1,343
Do you guys really think there will be more than 4 cores in a CCX?

A quick reminder of how the interconnect complexity grows.

The number of pathways between n objects is the combination C(n,2).

So for objects

2 = 1
3 = 3
4 = 6
5 = 10
6 = 15
7 = 21
8 = 28
16 = 120

One thing I've noticed about AMD is they don't give up on a design easily. Anything other than 4 core CCX would basically be a new chip.
Idk, everything I've read about Rome (EPYC Zen 2) says it has up to 48 cores. That's sampling now. 48 cores points to a 6 core ccx for Zen 2.
 

schmide

Limp Gawd
Joined
Jul 22, 2008
Messages
366
Idk, everything I've read about Rome (EPYC Zen 2) says it has up to 48 cores. That's sampling now. 48 cores points to a 6 core ccx for Zen 2.

48 cores can easily be done with 3 ccx per die 4 dies.

4 x 3 x 4 = 48
 

N4CR

Supreme [H]ardness
Joined
Oct 17, 2011
Messages
4,948
False rumor? They seem to obtain the 10--15% IPC from some ancient slide about performance

View attachment 90665

And the 8-core CCX rumor is also false. How in hell do you expect AMD to throw twice more cores and bigger cores in the same socket, whereas feeding all the cores with a dual channel?
I hate agreeing with you but you're partially right. No way in hell to get 10-15% IPC. With more IMC optimisation the ram will give gains like what we saw with Zen+ already.
8ccx is required, it may give some benefit too. And how else are they releasing 64 core same socket early next year? 8 core ccx is a given no brainer at this point.

7nm gives a considerable shrink in density. The extra IF links can be easily managed on the interposer, e.g. 65nm, so they will not take much die area. The only limit will be ram as you said. They already discussed this interposer approach would be needed for future products.
Conveniently this approach is leaving performance room on the table for the next socket with 8+ cores, ddr5 and 7nm+

I don't know how people in this thread are not aware AMD has already said, that they have 64 cores coming soon.
 
Last edited:

tayunz

[H]ard DCoTM October 2018
Joined
Sep 27, 2004
Messages
1,343
I hate agreeing with you but you're partially right. No way in hell to get 10-15% IPC. With more IMC optimisation the ram will give gains like what we saw with Zen+ already.
8ccx is required, it may give some benefit too. And how else are they releasing 64 core same socket early next year? 8 core ccx is a given no brainer at this point.

7nm gives a considerable shrink in density. The extra IF links can be easily managed on the interposer, e.g. 65nm, so they will not take much die area. The only limit will be ram as you said. They already discussed this interposer approach would be needed for future products.
Conveniently this approach is leaving performance room on the table for the next socket with 8+ cores, ddr5 and 7nm+

I don't know how people in this thread are not aware AMD has already said, that they have 64 cores coming soon.
Pretty sure 64 cores is Zen 3 though. I think it was a misquoted thing that got spread like the internet does.
 

gigaxtreme1

2[H]4U
Joined
Oct 1, 2002
Messages
3,577
Starship was the codename given to the 64 core part originally and the only way to do it on the platform frame is 8core ccx. LTCTL would be the linear interconnect
 
Last edited:

juanrga

2[H]4U
Joined
Feb 22, 2017
Messages
2,804
One memory channel per CCX. 2x 4-core CCX in Ryzen = 2 channel memory. 4x CCX in Threadripper = quad-channel memory
Core count per ccx does not dictate the number of channels it needs. Granted caching would need to be expanded to handle the higher core load, but that is well within the realm of possibility. I would think the higher level of multiplexing would cause some headaches but, well they have engineers to figure that one out.

If you double the number of cores plus increase the IPC plus increase the frequency, you need a lot more memory bandwidth to feed the cores. AMD cannot increase DRAM speed to 5000MHz, so it would have to increase the L3 per core to keep cores adequately feed. That increase on L3 comes at cost of die size together with the increases required by the IPC increase and the extra cores, but AMD is tied by the AM4 socket size.
 

juanrga

2[H]4U
Joined
Feb 22, 2017
Messages
2,804
My expectation for Zen2 are


I have always expected a 6-core CCX and two CCX per die, with the required memory bandwidth partially given by higher DRAM speed and partially by an increase in L3 cache per core. I am undecided about if Zen2 would include 256bit FMAC units and provide AVX512 compatibility by fusing both units.

The 6-core CCX improves performance by reducing the CCX-CCX communication and allows for 6-core Zen2 APUs.

Now imagine that there is no 6-core CCX, and Zen2 is 4-core CCX and 8-core die again ;)
 
Last edited:

gigaxtreme1

2[H]4U
Joined
Oct 1, 2002
Messages
3,577
You know, I have thought about the triangular interconnect to reduce latency. I'm inclined to agree with you architecturewise..
 

pendragon1

Extremely [H]
Joined
Oct 7, 2000
Messages
43,729
im gonna guess that someone will be wrong and try to back peddle and save face...
 

DuronBurgerMan

[H]ard|Gawd
Joined
Mar 13, 2017
Messages
1,340
10-15% IPC gain isn't happening. Maybe more like 5%. That would put it within spitting distance of Skylake IPC (within ~5%). That would be good enough.

No 8 core CCX, though I think a 6 core CCX is within the realm of possibility. Either that or an extra CCX per die. Contrary to some, though, I think the 6 core CCX is more likely. It would address more deficiencies in the design and would be the smarter play, IMHO.

Either way, I fully expect a 12 core mainstream part next year. AMD will want to maintain a core lead over Intel at every price band in order to maintain competitiveness.

I expect much higher boost clocks out of the 7nm process (high 4GHz range, say ~4.7 or 4.8), but base clocks won't be much (if any) higher, due to the need to maintain power/thermal limits on AM4 and TR4 with the core count increase, and in order to keep backward compatibility. In other words, I expect the efficiency gains of 7nm to be eaten up by the extra cores. But this should result in better overclocking headroom for folks who have the power deliver and cooling to deal with it.

I'm very excited for Zen 2. I think this is AMD's chance to push ahead of Intel on the mainstream. Right now, it's a wash IMHO between the 8700k and 2700X, depending on use. With a 12 core, better clocked, slightly higher IPC part, I think AMD takes the lead - even over an 8 core Intel part.
 

ccityinstaller

Supreme [H]ardness
Joined
Feb 23, 2007
Messages
4,241
I have a good board and good cooling so give me a 12c/24t part @ 4.5Ghz (I would love to see 4.7 but let's be real). Can't wait to use the 820mm of rad space I have lol. My 3 VEGAs live 99% of their life @ 1.3/1100Mhz -20% PL drawing under 135W a card lol. But they pay me soooo:p.
 

DuronBurgerMan

[H]ard|Gawd
Joined
Mar 13, 2017
Messages
1,340
I have a good board and good cooling so give me a 12c/24t part @ 4.5Ghz (I would love to see 4.7 but let's be real). Can't wait to use the 820mm of rad space I have lol. My 3 VEGAs live 99% of their life @ 1.3/1100Mhz -20% PL drawing under 135W a card lol. But they pay me soooo:p.

I expect 4.7 to be possible in boost clocks. And like the present situation, I expect max all core OC to be just shy of the max boost. OTOH, the power requirements will suck - so beefy VRMs will be needed. Base clock I expect to be high 3GHz range. Say, 3.7 like the 2700X, or maybe a touch higher. So more spread between base and boost clocks.
 

ccityinstaller

Supreme [H]ardness
Joined
Feb 23, 2007
Messages
4,241
I expect 4.7 to be possible in boost clocks. And like the present situation, I expect max all core OC to be just shy of the max boost. OTOH, the power requirements will suck - so beefy VRMs will be needed. Base clock I expect to be high 3GHz range. Say, 3.7 like the 2700X, or maybe a touch higher. So more spread between base and boost clocks.

I think you might see single core boost get to 4.7, but that does not intrest me *that* much...My cpu spends its time at 99% load 98% of the time, so max all core is what is important to me. If I were just a gamer, I would care a bit more about single core performance, but we have reached the point that 4.5Ghz is good enough to feed every GPU on the market @ 1440P above and 99% of 1080P users shy of some extreme 280hz CS player types.
 

Gideon

2[H]4U
Joined
Apr 13, 2006
Messages
3,151
Will have to see but so far all the rumors have been positive toward Zen 2, thus it may be more likely to have a larger performance increase over Zen which AMD stated was not as good as they hoped. So I think a 15% IPC increase may be possible, I am just less sure were clock speeds will land.
 

ccityinstaller

Supreme [H]ardness
Joined
Feb 23, 2007
Messages
4,241
Will have to see but so far all the rumors have been positive toward Zen 2, thus it may be more likely to have a larger performance increase over Zen which AMD stated was not as good as they hoped. So I think a 15% IPC increase may be possible, I am just less sure were clock speeds will land.

15% is a LOT. I think even 15% with the clockspeed increase would be a massive success and a good target for them to aim at to be honest. They had some low hanging fruit with the cache tweaks to Ryzen+, but it all depends on how much of a departure Zen2 is from the base design. I think 15% would be reasonable counting everything, and 10% would be a success.
 

DuronBurgerMan

[H]ard|Gawd
Joined
Mar 13, 2017
Messages
1,340
Will have to see but so far all the rumors have been positive toward Zen 2, thus it may be more likely to have a larger performance increase over Zen which AMD stated was not as good as they hoped. So I think a 15% IPC increase may be possible, I am just less sure were clock speeds will land.

AMD never stated that Zen wasn't as good as they hoped. Quite the opposite. Their target was a 40% IPC uplift vs. Bulldozer, and they ended up with well over 50% (52% or 54% IIRC). Zen exceeded expectations. Zen+ was just some process tweaks and latency tweaks, combined with a better boosting algorithm. Zen 2 will almost certainly incorporate design tweaks that will grant a little more IPC, but 15% is very doubtful. I'd be happy with 5%. The difference between Zen+ and Skylake in IPC is roughly 9%, excluding heavy AVX workloads. If AMD gets that difference down to a couple percentage points and makes up a little more of the clockspeed deficit, they are right where they should be, provided they are bringing extra cores to the table (which is likely).
 

DuronBurgerMan

[H]ard|Gawd
Joined
Mar 13, 2017
Messages
1,340
I have always expected a 6-core CCX and two CCX per die, with the required memory bandwidth partially given by higher DRAM speed and partially by an increase in L3 cache per core. I am undecided about if Zen2 would include 256bit FMAC units and provide AVX512 compatibility by fusing both units.

The 6-core CCX improves performance by reducing the CCX-CCX communication and allows for 6-core Zen2 APUs.

I agree with this, in general.
 
Top