New Zen microarchitecture details

[CentroX]

I wonder, what AMD will show on their AM4 boards at Computex. They will be there (the boards).

I thought computex was when AMDs polaris goes up against pascal? I wonder when we'll see summit ridge in action. If they want to meet their Q4 goal i hope atleast before summer.

 

[Abwx]

OTOH 8 Zen cores aren't comparable to the silicon area, leakage and power consumption of 8 32nm BD cores anymore.

At 14nm a 4M EXV would be something like 50W TDP, the only advantage Zen would get in respect of a 8M EXV is the ST perf, throughput wise the latter would be more than competitive, moreover given the L3 cache.

I wonder, what AMD will show on their AM4 boards at Computex. They will be there (the boards).

They will show...AM4 boards, since there will be BRidge as compatible SKU in the waiting of Zen...

 

[NTMBK]

At 14nm a 4M EXV would be something like 50W TDP, the only advantage Zen would get in respect of a 8M EXV is the ST perf, throughput wise the latter would be more than competitive, moreover given the L3 cache.

I think I'll take the decisions of AMD's engineers over random forum warrior speculation, thanks.

 

[Abwx]

I think I'll take the decisions of AMD's engineers over random forum warrior speculation, thanks.

That s the decision of the management, not of the enginers..

 

[el etro]

I think that Zen performance/efficiency may turn XV and any CON and even CAT core irrelevant. Don't matter on what node ZN is being fabbed.

 

[AtenRa]

The 35W TDP Server SKUs will be the one to watch from now on, those could be the next big thing in Servers the next 4-5 years.
I believe both AMD and Intel will have a nice fight over this market.

 

[CentroX]

I hope AM4 is going to be used for zen+ etc. I hate how intel always forced you to switch mobo because of their bazillion different sockets.

 

[Erenhardt]

It's foolish to love or hate a company.

When a poster talks of Intels "shitty" cpu upgrades, for example, one immediately looks to see what better cpu upgrades AMD has provided in recent years...

So when someone says after tasting ones dinner: "this tastes like shit", ones go out and lick dog's poo to deny him? And even if ones actually do this, it may be because he put a lot of heart in making this dish and he love it as is.

Why would one do the same when it comes to intel?

If we need to follow the aid of recent history, then Zen is yet another bulldozer derivative - improved carrizo. Can we not be silly?

 

[jpiniero]

Dual channel vs quadro? It is not really a big performance difference. Not on my X79 atleast.

Well, let me ask you this - why did you buy the 4820K over the 3770K? As you said the quad channel memory isn't all that useful on quad core. The extra PCIe lanes might be if you were doing SLI/CF.

 

[DrMrLordX]

The 35W TDP Server SKUs will be the one to watch from now on, those could be the next big thing in Servers the next 4-5 years.
I believe both AMD and Intel will have a nice fight over this market.

Intel has a big foot in that door thanks to Broadwell-D. It's a strong product.

 

[Dresdenboy]

They will show...AM4 boards, since there will be BRidge as compatible SKU in the waiting of Zen...

According to a rumour based on a MB vendor leak, AM4 boards are in production now. And Zen might possibly be shown at Computex, too. Would be nice, but we have to wait and see. Bristol Ridge already popped up in the Geekbench database.

 

[Doom2pro]

why their right minds would they design a separate die just with more cache ? It makes zero sense.

Regardless of what they did to the design, they MUST make large changes just to get it working on AM4 socket, they can't just take an existing die and slap it into an AM4 package... They clearly want a bridge to Zen on the new socket, so throwing in more cache doesn't seem that bad considering they need to tapeout a whole new chip anyway.

 

[Abwx]

According to a rumour based on a MB vendor leak, AM4 boards are in production now. And Zen might possibly be shown at Computex, too. Would be nice, but we have to wait and see. Bristol Ridge already popped up in the Geekbench database.

Hope so, they could eventually make a demo like they did with Polaris, with one or two hints about the perfs or perf/watt since only 4-5 months will be left before launch...

About the geekbench entry, are you talking of Stoney Ridge, i ve seen nothing about BR, is there a link.?.

 

[myocardia]

By the time it s realeased, and because it s released, the whole market will be shifting to lower prices within the pricey segments, currently a FX8370 is close to 200$ and should sail 50$ lower in 6 months, so there s no reason that a Zen will be more than 400$, that would be 2.66x the (downgraded) FX price.

I agree with this, except for the time period. I'm expecting it to take 8-9 more months from today, before any of us nobodies will be able to get our hands on a Zen. I am, however, expecting some of the people who are on a first name basis with AMD, like hardware review sites, along with very well-known forum dwellers, including the one has posted in this thread, to get one a bit sooner.

Beside there s no GPU, so the area shouldnt be bigger or significantly bigger than a 4790K for instance..

You think it will be somewhere in the 4790k's 177 mm2 size? I'm personally hoping that the Abu Dhabi foundry can bring it in under 200 mm2. I would consider ~200 mm2 a small victory, and anything at or under 180 mm2 to be a large victory. Here's to hoping that you are the one who is correct, since that would make it much more financially profitable for AMD.

 

[The Stilt]

After having a discussion about AM4 motherboards, it soon became obvious what kind of changes to the design the common socket and 14nm LPP process makes.

It is obvious that even the cheapest AM4 boards will be quite a lot more expensive than the cheapest AM3+ or FM2+ boards were. Unless AMD really starts to audit / enforce the design guidelines on AM4 motherboards, the consumers will be in a world of pain.

Zen appears to operate at very low voltage levels (even in relation to Intel 14nm parts) and because of that draw rather high amounts of current in relation to the power consumption. The increase in current draw is obvious as the process size goes down (lower voltage, P = I * V), however the actual increase appears to be a bit more than I personally anticipated. This means that the VRM(s) on the motherboards must be even tougher than on AM3+ motherboards. Because of the common socket (CPU & APU) the secondary VRM must be extremely beefy too. I would estimate the high-end boards should have 4 phase "SoC"-plane VRM (feeds the CNB and the iGPU) in order to be on the safe side. On AM3+ or FM2+ platforms 2-phase secondary VRM plane is the standard, even on high-end boards.

Naturally it is possible to make a Zen CPU targeted motherboards with restrictions applied on APU compatibility, in order to save on the cost. However a fully CPU & APU compatible high-end AM4 board will be expensive because of the VRM alone.

 

[mysticjbyrd]

So when someone says after tasting ones dinner: "this tastes like shit", ones go out and lick dog's poo to deny him? And even if ones actually do this, it may be because he put a lot of heart in making this dish and he love it as is.

Why would one do the same when it comes to intel?

If we need to follow the aid of recent history, then Zen is yet another bulldozer derivative - improved carrizo. Can we not be silly?

Glorifying a company and everything it does is different than loving/hating an individual product from a company.

 

[Doom2pro]

However a fully CPU & APU compatible high-end AM4 board will be expensive because of the VRM alone.

That's fine by me, I'd rather pay a little more than have to deal with possible shady/cheap Mobo designs that crap out and fry my hardware.

I'm not looking to buy a budget system when Zen arrives, I'm going ALL in, Zen + Polaris unless nVidia manages to pull a rabbit out of their hat (looking less and less likely).

 

[DrMrLordX]

I would estimate the high-end boards should have 4 phase "SoC"-plane VRM (feeds the CNB and the iGPU) in order to be on the safe side.

So we're looking at 8+4 phase designs? Sounds cool actually.

 

[The Stilt]

So we're looking at 8+4 phase designs? Sounds cool actually.

Certain manufacturers (the usual suspects...) will do 4+2 boards just as before. Will work on Zen CPUs as long as sufficient amount of airflow is provided to the VRM, but will work like a crap on Zen APUs (Raven Ridge) due the 5-6 fold current & power requirements on the SoC-power plane.

Despite there are two main power-planes, I expect the higher-end boards to use two VRM controllers, like FP4 mobile platform does. AM4 supports that kind of configuration and there are no single controller solutions which would support more than 2 secondary output phases.

There will be lot's of fooling around (cutting corners / cost) by the manufacturers, mark my words D:

 

[Abwx]

You think it will be somewhere in the 4790k's 177 mm2 size? I'm personally hoping that the Abu Dhabi foundry can bring it in under 200 mm2. I would consider ~200 mm2 a small victory, and anything at or under 180 mm2 to be a large victory. Here's to hoping that you are the one who is correct, since that would make it much more financially profitable for AMD.

Dont know exactly the size of the GPU in the 4790K but given that it has L3 cache the GPU size could allow for 4 other cores, so your 180-200mm2 seems to me a quite reasonable estimation...

Zen appears to operate at very low voltage levels (even in relation to Intel 14nm parts) and because of that draw rather high amounts of current in relation to the power consumption. The increase in current draw is obvious as the process size goes down (lower voltage, P = I * V),.

It s the other way around, the lower the voltage the lower the power, and at a fast rate since it s a...power law, that is a polynomial of degree 2..

The current I in function of voltage, frequency and capacitance is :

I = 2pi.V.F.C

Hence power is :

P = V.(2pi.V.F.C) = 2pi.(V^2)(F.C)

C is the capacitance of the circuit, in a CPU it s the sum of all switching transistors capacitances and is considered as being constant.

As the process goes down the capacitance goes down as well but the transistor (trans)conductance (gm) can be degraded, wich will require more voltage if the (trans)conductance loss is not compensated by a sufficently lowered capacitance as :

max Frequency = gm/C.

And gm is proportional to the square of the voltage, that s why increasing frequency by 1.21 require to increase voltage by sqrt(1.21) = 1.1.

Basicaly what was done with Intel s 14nm, or GF s 28nm for Carrizo, was to lower gm (to gain density with smaller transistors) and to decrease C accordingly so the net result is better efficency even if the voltage used is higher.

 

[RussianSensation]

Anyone else going all AMD when summit ridge launches? If summit ridge is better then my 4820K then i will switch teams. I am looking forward to go zen + polaris in 2016.

Bye bye intel, wont miss your shitty upgrades.

Alright then. I'll give you objective reasons why I decided to not wait for Zen.

1. I used FX9590 4.7-5Ghz as a point of reference. I then assumed that Zen will launch with 4.7-5Ghz clocks right out of the gate (I am being extremely favourable to Zen with this assumption). You will see later why I did this -- essentially I am creating the most favourable position for Zen.

2. Now I just need to compare how far behind FX9590 is against 6700K in terms of IPC to set a point of reference. This tells me just how far behind FX9590 is in terms of IPC. Relating back to point #1 above, by assuming Zen and FX9590 are clocked similarly, I can focus specifically on the gigantic gap that Zen would have to make up against Skylake to catch up.

Ready?

Video Card: NVIDIA GeForce GTX 980 Ti (6 GB / 384-bit GDDR5, 1000-1076 / 7010 MHz).

6700K is 103% faster

6700K is 64% faster

83% faster

50% faster

108% faster

Average gaming IPC delta: 6700K stock is beating FX9590 by 82%!
http://fcenter.ru/online/hardarticl...6700_Core_i5_6600_Core_i5_6500_i_Core_i5_6400

Wait for it, 6700K can also overclock another 400-600mhz from 4.2Ghz to 4.6-4.8Ghz. If we use a 4.6Ghz overclock, that's another 10% increase in IPC bringing us to 82% x 1.10 = 90% faster single threaded performance per core in games.

AMD is aiming to bring 40% increase in IPC. That means even if Zen comes out swinging with 4.7-5Ghz clocks, has another 10% overclocking headroom and has 40% increase in IPC, it's still nowhere close to Skylake.

Now I get it, no one games at 1280x800 with a 6700K. The point here is to show a CPU limited scenario of what would happen if Zen and Skylake were in a hypothetical CPU limited section of a game.

So what are the chances Zen has 4.7-5Ghz clocks, and another 10% overclocking headroom that i7-6700K has? All of a sudden it means that if Zen is clocked well below 4.7-5Ghz, you are now going to need to add 90% IPC advantage of Skylake + overclocking headroom of Zen to just catch up to 4.7-5Ghz of FX9590. See how dire the situation gets for Zen?

The next logical rebuttal is that Zen will offer more cores than mainstream Skylake in exchange for worse IPC. Since Zen will launch before SKL-E, we should instead compare it to BW-E's IPC. OK, that's covered too because BW-E's IPC is very close to Skylake's.

http://fcenter.ru/online/hardarticl..._Bridge_do_Skylake_Sravnitel_noe_testirovanie

http://www.hardware.fr/articles/940-6/cpu-sandy-bridge-vs-ivy-bridge-vs-haswell-vs-skylake-4-g.html

Further, Intel is about to introduce a 10-core BW-E and possibly drop prices on either the 6- or 8-core BW-E.

Bottom line: I think people have to set realistic expectations for Zen and I am not seeing that. Some are calling Zen a failure if it doesn't match or beat BW-E at minimum. Talk about living in dream land. Even Lisa Su has acknowledged in interviews that Zen is a start to a new path towards turning the firm around but she repeated that it's not a 1 time bullet/shot that will suddenly regain AMD CPU leadership position. It's rather a new fundamental base from which to build upon for Zen+ and so on.

See, people make a huge mistake of missing something. On the Intel side, we compare increase in IPC from SB to Skylake without realizing that FX9590's IPC is worse than Nehalems! Nehalem itself is a good 15-17% behind Sandy.

Proof:
Intel Core i7-870, 4C/4T, @ 2,80 GHz, 45 nm = 100%
Intel Core i7-930, 4C/4T, 2,80 GHz, 45 nm = 100%
vs.
Intel Core i5-2500K, 4C/4T, @ 2,80 GHz, 32 nm = 115%
Intel Core i7-2600K, 4C/4T, @ 2,80 GHz, 32 nm = 117%
http://www.computerbase.de/2011-01/test-intel-sandy-bridge/46/

Now we need to go way back to Nehalem era vs. FX8150 or compare FX8150 to Sandy to figure out just how far behind Bulldozer is.

AMD FX-8150, 4M / 8T, 3.60 GHz, 32 nm, Turbo = 100%
Intel Core i7-2600K, 4C / 8T, 3.40 GHz, 32 nm, Turbo, SMT = 128%
http://www.computerbase.de/2011-10/test-amd-bulldozer/14/

That means out of the 40% increase in IPC AMD is aiming for, almost 75% of that (30/40) will have to be used just to catch up to Sandy Bridge. This means we'd be lucky if Zen's IPC slots as a mid-point between IVB and Haswell.

So with that out of the way, where would an objective consumer stand right now?

Considering Intel CPUs hardly lose $ during resale, it's safe for 99% of consumers to buy Skylake now and see where Zen lands. It makes no sense to postpone an upgrade and wait 9+ months from now until Zen launches given the IPC characteristics of Broadwell and Skylake. Yet, once again I see Zen is being set up for failure from overhyping and having completely unrealistic expectations on this forum.

Basically, the way I see it, AMD's Zen strategy will be to offer more cores and price them more aggressively, coupled with cheaper motherboard prices. Offer 8 cores for the price of an i7 6700K and 4 cores for the price of an i3. The difference is Zen should have closer single-threaded performance, way lower power consumption, modern motherboard features and an upgrade path to Zen+ on the same socket/mobo. This will make it less trade-off of the Bulldozer strategy, but it still won't be enough to overcome Skylake's single threaded prowess. But, realistically speaking, the gap in IPC against modern Intel architectures is HUGE. If Zen nets 40%+ increase in IPC, with Intel's next major architecture out in 2018 with Icelake, this is where AMD can try to claw back another 10-15% IPC and slowly start catching up. I hope I am wrong as I want more competition but benchmarks of Skylake and IPC comparisons to FX9590 paint a different picture that I cannot just dismiss.

If my estimates of IPC of modern architectures are way off, then I apologize as I am open to reviewing other reviews/benches. Obviously, DX12 could save Zen's lower IPC by focusing on leveraging more cores but this is a far shot as there are too few DX12 games and chances are they won't be enough out in the next 12 months to suddenly start benefiting from 8-16 cores/threads to hide Zen's IPC deficits. Making matters worse, by the time Zen launches, Intel will have i7-7700K.

 

[The Stilt]

Dont know exactly the size of the GPU in the 4790K but given that it has L3 cache the GPU size could allow for 4 other cores, so your 180-200mm2 seems to me a quite reasonable estimation...




It s the other way around, the lower the voltage the lower the power, and at a fast rate since it s a...power law, that is a polynomial of degree 2..

The current I in function of voltage, frequency and capacitance is :

I = 2pi.V.F.C

Hence power is :

P = V.(2pi.V.F.C) = 2pi.(V^2)(F.C)

C is the capacitance of the circuit, in a CPU it s the sum of all switching transistors capacitances and is considered as being constant.

As the process goes down the capacitance goes down as well but the transistor (trans)conductance (gm) can be degraded, wich will require more voltage if the (trans)conductance loss is not compensated by a sufficently lowered capacitance as :

max Frequency = gm/C.

And gm is proportional to the square of the voltage, that s why increasing frequency by 1.21 require to increase voltage by sqrt(1.21) = 1.1.

Basicaly what was done with Intel s 14nm, or GF s 28nm for Carrizo, was to lower gm (to gain density with smaller transistors) and to decrease C accordingly so the net result is better efficency even if the voltage used is higher.

You missed the whole point.

95W @ 1.0V = 95A
95W @ 1.3V = 73A

30% or 22A is a huge amount of current and can easily require one or two additional VRM phases, depending on the component quality (cost).

 

[mysticjbyrd]

Alright then. I'll give you objective reasons why I decided to not wait for Zen.

1. I used FX9590 4.7-5Ghz as a point of reference. I then assumed that Zen will launch with 4.7-5Ghz clocks right out of the gate (I am being extremely favourable to Zen with this assumption). You will see later why I did this -- essentially I am creating the most favourable position for Zen.

2. Now I just need to compare how far behind FX9590 is against 6700K in terms of IPC to set a point of reference. This tells me just how far behind FX9590 is in terms of IPC. Relating back to point #1 above, by assuming Zen and FX9590 are clocked similarly, I can focus specifically on the gigantic gap that Zen would have to make up against Skylake to catch up.

Ready?

Usually you have great posts, but this is utter nonsense.

Some cherry picked game benchmarks seems like a really poor metric to base a conclusion on with regards to the potential of Zen. It's a completely different architecture, which is mirroring intel. You are comparing apples and oranges in numerous ways.

I doubt Zen will match Skylake in most games, unless they can really make use of all the extra threads, which is doubtful for the near future. However, the difference in game performance shouldn't be huge.


I think a better, and far more simple comparison, would be just to look at 5820k benchmarks.

 

[deasd]

Average gaming IPC delta

Cough......sorry I can't even teach every single one on common knowledge of system architecture and programming.
Gaming is gaming, IPC is IPC, the former is used to test WHOLE platform performance(include GPU, system bus, memory, I/O), latter is just a unit of measurement, it is for programmer to measure some instruction latency. How can you test so-call IPC(not to mention you don't know what instruction you test) in a graphic/memory/I/O stressing circumstance? What tools had you used??
It seems so many people cannot figure out yet...... but I won't worry just wait things turn out.

 

[RussianSensation]

Usually you have great posts, but this is utter nonsense.

Some cherry picked game benchmarks seems like a really poor metric to base a conclusion on with regards to the potential of Zen. It's a completely different architecture, which is mirroring intel. You are comparing apples and oranges in numerous ways.

How? I am comparing CPUs how they have been compared for ages, going back 20 years. In that link they have non-gaming tests but that's not the point since many of us game. Point is, FX9590 is miles behind in single core performance. Using FX9590 as a point of reference + adding 40% single threaded performance isn't enough to catch up to Skylake. My analysis backs it up. Many, many people on this forum already concluded what I did months ago.

I doubt Zen will match Skylake in most games, unless they can really make use of all the extra threads, which is doubtful for the near future. However, the difference in game performance shouldn't be huge.

How do you know that? FPS doesn't tell the full story.

[/IMG]

i7 6700K demolishes the 2600K in games.
https://www.youtube.com/watch?v=VDo-j00vUtw

The main reason gamers with 2600K haven't upgraded is they probably have it overclocked to 4.6-4.9Ghz. However, if someone is making a new buying decision, they'd never pick 2600K @ 4.7Ghz over a stock i7 6700K. If Zen barely gets to IVB IPC, that's exactly the position it'll be in unless AMD suddenly pushes 6-8 cores for the price of Intel's 4.

I think a better, and far more simple comparison, would be just to look at 5820k benchmarks.

That's an irrelevant comparison since by the time Zen launches, we'll have 6820K Broadwell-E. I linked BW's single threaded IPC already and it smashes Sandy which smashes Nehalem, which smashes FX9590's architecture.

Cough......sorry I can't even teach every single one on common knowledge of system architecture and programming.
Gaming is gaming, IPC is IPC, the former is used to test WHOLE platform performance(include GPU, system bus, memory, I/O), latter is just a unit of measurement, it is for programmer to measure some instruction latency. How can you test so-call IPC(not to mention you don't know what instruction you test) in a graphic/memory/I/O stressing circumstance? What tools had you used??
It seems so many people cannot figure out yet...... but I won't worry just wait things turn out.

As a consumer, I don't need to know anything about programming. I look at a test suite with 20 applications/programs/games and see how Processor 1 @ XYZ clocks compares to Processor 2 @ XYZ clocks. Processor 1 (Sandy) is faster than Processor 2 by 28%. That tells me, across an average of programs I may run, Processor 1 will be faster by 28%.

You can get technical and discuss semantics of the term IPC, fine, call it single threaded/single core performance. I have already shown that Sandy Bridge's single core performance is 28% faster than Bulldozer. In turn, Skylake is at least 25% faster than Sandy bridge in IPC/single threaded per core performance. That's being generous too given how much of a cow FX9590 is for games in CPU limited scenarios.

You didn't even bother reading the reviews/charts I linked. Focusing in on the definition of IPC changes little about the actual data.

I can link many reviews that highlight the point I am making.

AMD needs a 70-80%+ increase in single-threaded performance/IPC to catch up to a Skylake core. They are bringing 40%. Most consumers are not going out and buying a 5820K over an i5-6600K/i7-6700K despite that CPU having a far smaller deficit in IPC to Skylake than Zen will. You can connect the dots. The market values faster single core CPUs with lower power usage, not MOAR cores with lower IPC.

In simpler terms, if Zen brings 5820K performance for $350, Intel's workstation platform will still be better with a 6820K, while the mainstream will have faster quad-core i7-7700K. Where does that leave Zen? MOAR cores for less $? Ya, that already failed over the last 5 years.