[Austin Statesman] AMD sees a way forward (with new Zen design)

[frozentundra123456]

Zen will be an 8C/16T part at the top. There are plenty of unknowns though. Will AMD hit its perf/clock goals? What kind of clocks will AMD hit? What will AMD charge for the thing?

The hype train is running full throttle, and with every new thread the expectations get more and more outlandish. We have gone from a somewhat realistic estimate of SB ipc to now being faster than Skylake. Before it comes out we will have AMD fans predicting it will provide cold fusion, cure cancer, and solve global warming.

 

[Arachnotronic]

The hype train is running full throttle, and with every new thread the expectations get more and more outlandish. We have gone from a somewhat realistic estimate of SB ipc to now being faster than Skylake. Before it comes out we will have AMD fans predicting it will provide cold fusion, cure cancer, and solve global warming.

I think the fact that AMD will be sticking 8 of these CPUs into a 95W thermal envelope is a dead giveaway that it's not going to clock very high.

 

[Sweepr]

The hype train is running full throttle, and with every new thread the expectations get more and more outlandish. We have gone from a somewhat realistic estimate of SB ipc to now being faster than Skylake. Before it comes out we will have AMD fans predicting it will provide cold fusion, cure cancer, and solve global warming.

Currently the hype train is at better than Haswell IPC, 8-cores at 95W TDP, high clocks (according to some this 14nm is better than Intel's), mainstream prices and 2016 launch without further delays.

And don't dare to question any of the above, otherwise you will be called a fanboy and there will be trolls asking for sources.

 

[Boze]

The hype train is running full throttle, and with every new thread the expectations get more and more outlandish. We have gone from a somewhat realistic estimate of SB ipc to now being faster than Skylake. Before it comes out we will have AMD fans predicting it will provide cold fusion, cure cancer, and solve global warming.

Let's try to be somewhat realistic... Zen will be used to power the real life Starship Enterprise that AMD will be building in 2018.

 

[AtenRa]

Le sigh.

Intel isn't going anywhere. Apple will not kill them. Zen won't kill them either.

But here's one thing it might do . . .

http://cdn.overclock.net/0/05/052237f1_CZ-CB10-3.4G-Static.png

From:

http://www.overclock.net/t/1560230/jagatreview-hands-on-amd-fx-8800p-carrizo/400_100#post_24310470

And the Carrizo in question was running @ 3.4 GHz (static no throttling) during the Cinebench R10 run:

http://www.overclock.net/t/1560230/jagatreview-hands-on-amd-fx-8800p-carrizo/400_100#post_24314113

So what can we tell from a Cinebench R10 score of 13146 from a 2m/4t Excavator @ 3.4 GHz?

13146 * 1.4 * 4 = 73617.6 (round to 73618)

That's assuming 8c/16t Zen, 40% increase in IPC, and a clockspeed of 3.4 GHz. Now let's look at Skylake running at a static clockspeed. How about 4.8 GHz Skylake:

http://www.overclockers.com/wp-content/uploads/2015/07/intel_i76700K-46.jpg

From:

http://www.overclockers.com/intel-skylake-i7-6700k-cpu-review/

40731. Nice score right? Assuming an 8c/16t Skylake-E could hit the same clockspeed, you'd be looking at a score of 81462.

Normalizing for clockspeed, Skylake has a performance of ~2121 CB per GHZ per core. Kabylake should be no different. Zen (according to AMD's estimate of a 40% IPC increase) should have a performance of ~2707 per GHz per core.

Hmmmmm!

Those are not single thread scores.

 

[R0H1T]

Even a blind chicken finds the odd kernel of corn.

Yeah like the blue brigade foretelling the end of times for AMD, ARM, Apple for the better part of this decade?

And what, pray tell, is so "egregious" about Intel's margins?

Well for one they expect everyone to pay the Wintel tax even when Windows' marketshare is being corroded to the point that MS will be a non player in anything that doesn't sit on your desk come the next decade. The server monopoly is also a happy of consequence of Wintel duopoly (as IBM left the field), again the process lead they have is being eroded & against a horde of chipmakers that may or may not be aiming for (their own?) servers I see Intel ~

a) becoming a shadow of their current self if they don't start lowering their margins eventually because the largest tech firms all have within them the ability to pull something like that off.

b) accept the fate of low(er) margins as the new normal & stay relevant till they carve a niche into something like IoT, the latter may or may not be the next big bubble waiting to burst.

The electronics industry as a whole has moved towards "disposable" products, to put it simply consumers expect more perf for their $ (even though the performance is stagnating) & are not going to pay a premium just for performance. As of today I could get a 2GB (RAM) LTE smartphone with Snapdragon 410 for as little as $85 off contract, in a fire sale. The reason Apple gets a premium is because they bundle their software ecosystem with such devices, iOS & OS X both command a premium due to the Apple experience. So unless Intel gets another breakthrough, like graphene or SiGe, come 7nm ARM will be (too) close to them in the server arena in terms of relative performance & they don't suffer from the same perf/W ailment that has hurt AMD in the past. It just takes one to set an example & then we'll see the floodgates open.

Anyone not having their own software or services ecosystem in the coming decade is ill positioned as the electronics (& semiconductor) industry as a whole is moving towards low margin, high volume products.

 

[Arachnotronic]

Yeah like the blue brigade foretelling the end of times for AMD, ARM, Apple for the better part of this decade?Well for one they expect everyone to pay the Wintel tax even when Windows' marketshare is being corroded to the point that MS will be a non player in anything that doesn't sit on your desk come the next decade. The server monopoly is also a happy of consequence of Wintel duopoly (as IBM left the field), again the process lead they have is being eroded & against a horde of chipmakers that may or may not be aiming for (their own?) servers I see Intel ~

a) becoming a shadow of their current self if they don't start lowering their margins eventually because the largest tech firms all have within them the ability to pull something like that off.

b) accept the fate of low(er) margins as the new normal & stay relevant till they carve a niche into something like IoT, the latter may or may not be the next big bubble waiting to burst.

The electronics industry as a whole has moved towards "disposable" products, to put it simply consumers expect more perf for their $ (even though the performance is stagnating) & are not going to pay a premium just for performance. As of today I could get a 2GB (RAM) LTE smartphone with Snapdragon 410 for as little as $85 off contract, in a fire sale. The reason Apple gets a premium is because they bundle their software ecosystem with such devices, iOS & OS X both command a premium due to the Apple experience. So unless Intel gets another breakthrough, like graphene or SiGe, come 7nm ARM will be (too) close to them in the server arena in terms of relative performance & they don't suffer from the same perf/W ailment that has hurt AMD in the past. It just takes one to set an example & then we'll see the floodgates open.

Anyone not having their own software or services ecosystem in the coming decade is ill positioned as the electronics (& semiconductor) industry as a whole is moving towards low margin, high volume products.

R0H1T, you're jumping ahead of yourself here. You have not shown that Intel's margins are "obscene." Please provide a proof of this claim and once this claim is established, we can then start discussing the potential ramifications of these "obscene margins" disappearing

 

[R0H1T]

R0H1T, you're jumping ahead of yourself here. You have not shown that Intel's margins are "obscene." Please provide a proof of this claim and once this claim is established, we can then start discussing the potential ramifications of these "obscene margins" disappearing

I guess that's what you're calling it, anyway perhaps I didn't make it clear enough -> their server margins are insane, as compared to anything else they sell, & you know that just as well.

 

[Arachnotronic]

I guess that's what you're calling it, anyway perhaps I didn't make it clear enough -> their server margins are insane, as compared to anything else they sell, & you know that just as well.

OK, we have found common ground in "Intel's server chip margins are very high."

That said, gross profit margins in businesses like server chips (which are fairly low volume) are typically quite high. This is due to the fact that the volumes in this market are quite low and the R&D intensity to develop these chips/solutions are very high.

In a nutshell, it would be very hard for a company to even break-even on server/data-center type chips without commanding high gross margins. There's really nothing "obscene" about the types of margins Intel sees in its data center biz, IMO.

 

[AtenRa]

The hype train is running full throttle, and with every new thread the expectations get more and more outlandish. We have gone from a somewhat realistic estimate of SB ipc to now being faster than Skylake. Before it comes out we will have AMD fans predicting it will provide cold fusion, cure cancer, and solve global warming.

IF (Again double capital and bold) ΖΕΝ IPC is 40% higher than Excavator then it is easy to estimate a close to Haswell IPC.

Take Cinebench 15 (Intel optimized to the teeth)
A8-7600 Cinebench single core score is 88
Add another 5% for the Excavator and we get to 92
Now add 40% for the ZEN and we get 129

Haswell Core i3 score is 137

This is an estimation,it could be wrong it could be close to what Zen will be. The real question is, what clocks will they be able to have with a 6C 12T SKU at 95W TDP. If they can reach close to 4GHz then they will be very competitive against Broadwell-E in late 2016 early 2017.

 

[Dresdenboy]

Guys, what's happening between some of you here is provocation by one side, which is in the comfortable seat of following a market dominator like a strong big brother who'll beat up anyone not liked by the little one. The other side reacts to the provocation, and some leave the path of careful estimations as they want to counter with more than what's publicly known about Zen. That's the pattern.

DrMrLordX used "might" and "assumed" which is not hyping but just speculation. Others who want to provoke take the numbers as a fact. Why?

 

[R0H1T]

OK, we have found common ground in "Intel's server chip margins are very high."

That said, gross profit margins in businesses like server chips (which are fairly low volume) are typically quite high. This is due to the fact that the volumes in this market are quite low and the R&D intensity to develop these chips/solutions are very high.

In a nutshell, it would be very hard for a company to even break-even on server/data-center type chips without commanding high gross margins. There's really nothing "obscene" about the types of margins Intel sees in its data center biz, IMO.

That's the thing, so what happens when Intel hits the wall called physics? Their superior process lead will vanish, ARM with its cheaper alternatives will be a great way for many, like Apple/Google et al, to save money on the most expensive part of their operations (second most expensive for Apple) i.e. hardware.

The basis of this hypothesis is that even the likes Google & FB will reach a level of market saturation, just like MS, beyond which their growth will stagnate or deflate. The only way they could continue to expand is by spending tons on R&D, part of that exercise will likely include building their own hardware. The key to all of this is vision, as I've emphasized. If it's (custom ARM) good enough for 90% of your workload then it makes no sense to not pursue such an endeavor yourself, applies to Apple & Samsung currently, cause the $$ savings will be massive. Then comes the outlier, Mediatek & other Chinese companies operate on low margin & high volume business. They could just as well make server chips, spending billions in the process, potentially doubling their margins & still be cheaper than Intel in the targeted market segment.

It does look like all I'm predicting is doom & gloom for Intel, however I've also outlined what they can do to arrest this growing tide i.e. accept lower margins or make another tech breakthrough to continue charging a premium, especially for servers.

 

[R0H1T]

This is an estimation,it could be wrong it could be close to what Zen will be. The real question is, what clocks will they be able to have with a 6C 12T SKU at 95W TDP. If they can reach close to 4GHz then they will be very competitive against Broadwell-E in late 2016 early 2017.

I'd say 3~3.5Ghz (turboed) is more realistic at 95W TDP, considering this is first gen Zen & AMD's first attempt at 14nm FF, clocks & IPC should get higher with the passage of time. Btw is there any chance (desktop) Excavator will be on 14nm FF anytime soon, if at all?

 

[AtenRa]

I'd say 3~3.5Ghz (turboed) is more realistic at 95W TDP, considering this is first gen Zen & AMD's first attempt at 14nm FF, clocks & IPC should get higher with the passage of time.

Im expecting a 6C 12T Broadwell-E at close to 4GHz in Q2-Q3 2016. If ZEN will not be able to reach close to the same performance they will have to go for better perf/watt. But im expecting ZEN will be able to clock to 4GHz at 6C 12T with a 95W TDP for the desktop part.

Btw is there any chance (desktop) Excavator will be on 14nm FF anytime soon, if at all?

There is a rumor for a 14nm desktop Excavator APU for 2016. We will have to wait and see.

 

[R0H1T]

Im expecting a 6C 12T Broadwell-E at close to 4GHz in Q2-Q3 2016. If ZEN will not be able to reach close to the same performance they will have to go for better perf/watt. But im expecting ZEN will be able to clock to 4GHz at 6C 12T with a 95W TDP for the desktop part.

So there'll be multiple SKU's (possibly smaller dies) with Zen?

There is a rumor for a 14nm desktop Excavator APU for 2016. We will have to wait and see.

The construction cores should do well going forward, especially on 14nm FF, we'll only see how far Intel is ahead of AMD when they're on a similar process node, architectural differences aside.

 

[AtenRa]

So there'll be multiple SKU's (possibly smaller dies) with Zen?

I dont believe so, they most probable will only use a single die for all the Desktop/Workstation and Server parts.
What i mean is that you dont need the highest perf/watt for the Desktop parts. You could have a 4GHz 6C 12T at 95W model for the Desktop/Workstation and a 3.0GHz 6C 12T at 65W Server model.

The construction cores should do well going forward, especially on 14nm FF, we'll only see how far Intel is ahead of AMD when they're on a similar process node, architectural differences aside.

An Excavator APU at 14nm FF will be amazing, it could really be the attraction of 2016. You could have the same performance of Godavari at half the TDP and half the die size. Including DDR-4 memory (3000MHz or above) will make the iGPU performance increase by 20-30% or even higher with any architectural enhancements (GCN 1.2).
That will make it faster than Broadwell-C GT3e and way cheaper (no eDRAM or HBM etc). This could be the first 45-65W TDP APU truly 1080p Gaming capable at sub $150 mark.

 

[cbn]

There is a rumor for a 14nm desktop Excavator APU for 2016. We will have to wait and see.

Link?

 

[burninatortech4]

I dont believe so, they most probable will only use a single die for all the Desktop/Workstation and Server parts.
What i mean is that you dont need the highest perf/watt for the Desktop parts. You could have a 4GHz 6C 12T at 95W model for the Desktop/Workstation and a 3.0GHz 6C 12T at 65W Server model.



An Excavator APU at 14nm FF will be amazing, it could really be the attraction of 2016. You could have the same performance of Godavari at half the TDP and half the die size. Including DDR-4 memory (3000MHz or above) will make the iGPU performance increase by 20-30% or even higher with any architectural enhancements (GCN 1.2).
That will make it faster than Broadwell-C GT3e and way cheaper (no eDRAM or HBM etc). This could be the first 45-65W TDP APU truly 1080p Gaming capable at sub $150 mark.

I'm assuming this would be on the new AM4 socket and not FM2+? I would really like one last chip on my motherboard before a system overhaul.

 

[AtenRa]

Link?

If i remember correctly it was from a linkedin account, I cant find the link now.

 

[AtenRa]

I'm assuming this would be on the new AM4 socket and not FM2+? I would really like one last chip on my motherboard before a system overhaul.

Yes 2016 is AM4.

 

[DrMrLordX]

Again you're fixated on Cinebench, when I've said once already that I just drew a name out of a hat...

Then pick a different benchmark. Currently I have reliable Excavator numbers for Cinebench R10, y-cruncher (AVX, AVX2, SSE3, and xOP), 3DPMRedux and Linpack (patched for AMD CPUs so that it'll run). I might be able to convince The Stilt to run some more tests, though I'm not sure what is the status of his Carrizo test machine.

I'll give you the same deal I offered What'sHisFace. If a $370 Zen can compete against a $370 Core i7, and if a $600 Zen can compete against a $600 Skylake-E, and if a $1000 Zen can compete against an i7 6960X or i7 7960X, then I'll never post here again.

Take the deal. If you're convinced that AMD will release some magical savior product. If not, keep blowing smoke.

That's not much of a deal, since I get nothing from it.

The hype train is running full throttle, and with every new thread the expectations get more and more outlandish. We have gone from a somewhat realistic estimate of SB ipc to now being faster than Skylake. Before it comes out we will have AMD fans predicting it will provide cold fusion, cure cancer, and solve global warming.

No. AMD said from the start:

1). 8c/16t SMT-based uarch, TDP target 95w
2). 40% better IPC than Excavator

And that's it. Try to avoid hyperbole.

I think the fact that AMD will be sticking 8 of these CPUs into a 95W thermal envelope is a dead giveaway that it's not going to clock very high.

The pessimist in me is guessing 3 GHz for a base clock, with some kind of turbo. Who knows, maybe they'll do better. Carrizo with HDL can still hit 3.4 GHz (maybe higher?) so there's some hope for better clocks on the 14nm LPP process.

Those are not single thread scores.

No, they are not. I do the best I can with the available data. I would expect single-threaded scores to produce higher per-GHz numbers for both architectures since the numbers I used have imperfect MP/multicore scaling. Which is inevitable.

 

[DrMrLordX]

IF
Add another 5% for the Excavator and we get to 92

Interestingly enough, several of the benchmarks I got The Stilt to run showed IPC improvements of greater than 5% over Kaveri.

xOP y-cruncher was 11% faster, Linpack was 11% faster, and 3DPMRedux was a whopping 21% faster. However, Cinebench R10 was only 5% faster, so your numbers are not so unreasonable.

Guh, double post. Oh well.

 

[mrmt]

I guess that's what you're calling it, anyway perhaps I didn't make it clear enough -> their server margins are insane, as compared to anything else they sell, & you know that just as well.

Do you know what are the margins of IBM Mainframe contracts? I'm talking about equipment only, with services I won't even comment.

Server SKUs are inherently low volume because it needs a lot of extra, specific features, components and R&D over consumer parts to back them up, so you basically will end up with expensive chips per unit but not necessarily profitable business units (As IBM and SUN shows us). Whatever ARM IHV that comes to the server market will end up with a similar margin structure, even if they end up with an overall cost structure because of the larger ARM ecosystem.

 

[VirtualLarry]

An Excavator APU at 14nm FF will be amazing, it could really be the attraction of 2016. You could have the same performance of Godavari at half the TDP and half the die size. Including DDR-4 memory (3000MHz or above) will make the iGPU performance increase by 20-30% or even higher with any architectural enhancements (GCN 1.2).
That will make it faster than Broadwell-C GT3e and way cheaper (no eDRAM or HBM etc). This could be the first 45-65W TDP APU truly 1080p Gaming capable at sub $150 mark.

When you describe it like that... it sure sounds tasty! I hope that does get released.

 

[Dresdenboy]

If i remember correctly it was from a linkedin account, I cant find the link now.

There is one LinkedIn account mentioning 2017.

Interestingly enough, several of the benchmarks I got The Stilt to run showed IPC improvements of greater than 5% over Kaveri.

xOP y-cruncher was 11% faster, Linpack was 11% faster, and 3DPMRedux was a whopping 21% faster. However, Cinebench R10 was only 5% faster, so your numbers are not so unreasonable.

I even don't expect a big jump in Cinebench performance for Zen, as it already is rather cache friendly and raw FP throughput of XV is good. Of course a lower power consumption at iso clock frequency would allow for higher scores. Other FP heavy benchmarks might improve more than CB due to likely better data flow handling.


BTW, something to disprove:

GLOBALFOUNDRIES is yielding on its 14nm technology and is on schedule to support multiple product tape-outs and volume ramp in 2015.

2015 is over soon.