AMD zen pre-release thread!

[Vesku]

Well, that has nothing to do with the fact that even Bulldozer(FX8150) was faster than Core i7 920/860.
http://anandtech.com/bench/product/434?vs=108

Two years later, more power hungry on similar node, not quite equal in single thread which diminishes the gap in multiple threaded tasks.

Shame AMD hadn't made the Vishera improvements for their 2011 version. The 8350 would have been an easier sales pitch. Basically if they could have shifted their R&D ahead a 12-18 months they'd be doing OK but that's almost a state of being for AMD. In fact I think it was being behind that saved them from the initial pitfalls of following Intel's P4 design. But for whatever reason they stumbled back into it when they reached their same phase of development, "Hey guys lets make a chip designed to clock really high and with impressive theoretical throughput instead of reasonable clocks, balanced old and new instruction IPC, and fast cache. We'll call it BullBurst. No, no, NetDozer. BullDozer!"

http://www.anandtech.com/show/661/2

Why Dirk, why?

As i have said so many times, efficiency can be achieved at high performance as well. Efficiency doesnt mean low power and low performance only.
I dont expect 140-150W TDP CPUs but ZEN will not be a low power low cost small core design.

I expect it to be some sort of "big" core but most likely optimized for ~25-65W power envelope and whose efficiency focus will make producing ~100+W performance SKUs difficult. It's going to look a lot more like server desktop Broadwell than SandyBridge, basically.

Was $50 cheaper too

Yet $30 more than the i5 2500K that had been out for almost a year. Should have at least matched its pricing.

 

[AtenRa]

Two years later, more power hungry on similar node, not quite equal in single thread which diminishes the gap in multiple threaded tasks.

I dont disagree but i was only pointing out that even Bulldozer was faster than Core i7 920/860.

I expect it to be some sort of "big" core but most likely optimized for ~25-65W power envelope and whose efficiency focus will make producing ~100+W performance SKUs difficult. It's going to look a lot more like server desktop Broadwell than SandyBridge, basically.

Im expecting up to 95W TDP Desktop/server SKUs and higher TDP Server parts.

 

[mrmt]

if you listen to jim kellar, he talks about this form the perspective of not boiling the ocean. Seattle was to get arm to x86 server platform level, skybridge was to figure out how to create the interconnecting fabric allowing for either x86 or arm to drop into the SOC. He then specifically states that the fabric in K12 is significantly upgraded in terms of features and in terms of scope ( specifically mentions high performance servers).

Either Jim Keller quote is out of context or not true at all, or else AMD has a *huge* marketing problem on its way, because AMD is feeling the rest of the chain and their investors with information that Zen and K12 won't be targeted at Intel-type of high performance server.

 

[Abwx]

Your source is the words of a random internet tech journalist.

Let me tell you how this will play out. The people basing their predictions on actual AMD statements and documents will be right(shintai). The people whose predictions run contrary to AMDs statements will be wrong(you).

Your confirmation bias has caused you to believe a random person over official AMD slides.

What is random is your post, Techreport interviewed Mark Papermaster and Jim keller, that s the whole point of this article, so much for being biaised...

You should really learn to read before posting what amount to trolling given the sayings in the article.

A quote perhaps ?.:

AMD's execs noted that these are high-frequency, high-performance CPU cores that will span the range from laptops to desktops to servers—and not just "microservers."

Is the journalist talking from his own words or is he just repeating what he was told by thoses two engineers..?..

Are TR journalists also executives at AMD..?

These two CPU microarchitectures will be, in the words of CTO Mark Papermaster, "sister cores."

Yet another TR journalist opinion.?.

Be ready to get your pathetic prevision, not to say fud, bursted.

 

[Vesku]

Im expecting up to 95W TDP Desktop/server SKUs and higher TDP Server parts.

If they do it will be from core count blow out ala big Xeon not frequency. Maybe by late 2016 early 2017 the performance desktop market will be more enthused at sub $300 Intel 6+ core SKUs and consider an AMD 8-10 core that is a bit slower ST. Probably since Intel has almost gotten around to it with the 5820K at $389.

 

[AtenRa]

If they do it will be from core count blow out ala big Xeon not frequency. Maybe by late 2016 early 2017 the performance desktop market will be more enthused at sub $300 6+ core SKUs. Probably since Intel has almost gotten around to it with the 5820K at $389.

Dont forget those are only CPUs, they dont have any iGPUs. An 8 Core 16 Thread 95W TDP CPU at 14nm FF could be way powerful in 2016 even with a Sandy/Ivy single thread performance.

 

[coercitiv]

Either Jim Keller quote is out of context or not true at all, or else AMD has a *huge* marketing problem on its way, because AMD is feeling the rest of the chain and their investors with information that Zen and K12 won't be targeted at Intel-type of high performance server.

You can judge the context yourself, the video is here.

At 2:40 you can find the "boil the ocean" part.

as we do a new processor, it's not our first time doing high performance

There's a specific question related to differences between A57 and K12 at 11:08, which is answered by Jim Keller in the presence of several AMD officials, including Lisa Su:

They're differentiated in size, performance and power, and they're both beautiful parts, one will target the higher end and one will target the other end of the market.

Watch the entire video: they never compare the new cores to Intel's products, but Keller specifically mentions "high performance" and "high frequency" attributes in relation to the new cores.

 

[Abwx]

Watch the entire video: they never compare the new cores to Intel's products, but Keller specifically mentions "high performance" and "high frequency" attributes in relation to the new cores.

Yes, and what is your interpretation of "high performance".?.

Notice that he didnt said "High performance per watt"...

As to not relate it to Intel, you think that they are ignoring that this is their main competitor and that Zen/K12 will be in a collision course with said firm s Xeons.?.

Some people here should definitly give up replacing AMD s intentions by their own wet dreams of seeing this firm failing to compete with their prefered brand...

2016 AMD Server
-Up to 32Core/64Thread K12
-Up to 32Core/64Thread Zen

K12 Target Performance:2x~ Cortex-A57

Project SkyBridge on server
-Pin compatible,Common interface in SoC
-Same core size,same core shape
-Logical and Physical design level compatibility

2017 AMD HPC APU
-Next Generation FirePro and Opteron
-200~300W TDP

via PC Cluster Consortium on Japan

http://semiaccurate.com/forums/showpost.php?p=231022&postcount=987

 

[AtenRa]

200-300w TDP APU ???

 

[Abwx]

200-300w TDP APU ???

Yes it seems, but look to be something else than a DT APU :

http://semiaccurate.com/forums/showthread.php?t=8238&page=100

 

[Vesku]

200-300w TDP APU ???

Some CPU attached to a full FirePro GPU on one PCB. Same concept Nvidia is hinting at with Pascal+, ARM Cores integrated. Both are sort of catching up to Intel's Xeon Phi in that aspect running their own, presumably, Linux OS.

 

[coercitiv]

Yes, and what is your interpretation of "high performance".?.

Sometimes you're so bent on showing the world + dog that they're wrong, you don't even notice people who are at least partially agreeing with you.

 

[Vesku]

Keep in mind 4-5W per core is a different tier of performance compared to current ARM big core targets of ~1W per core (A57). So Keller's statements could easily be referring to an efficiency focused big core as we've been discussing rather than a king of the hill type of design. 4-5x W/core and ~4-5x the size would certainly be very powerful compared to A57.

 

[Idontcare]

Sometimes you're so bent on showing the world + dog that they're wrong, you don't even notice people who are at least partially agreeing with you.

Agreed.

 

[beginner99]

I dont disagree but i was only pointing out that even Bulldozer was faster than Core i7 920/860.

Im expecting up to 95W TDP Desktop/server SKUs and higher TDP Server parts.

Your own link shows that bulldozer loses to i7-860 in all single-threaded benches...and so does Piledriver. In fact they only heavy threaded benches and all of the are except Cine-Bench single-threaded. Of course the 920 loses more due to it's low clock. And I was usually following gaming benches from other sites where the i7-860 single-threaded performance was easily able to beat Piledriver in FPS. But that's actually off topic. I still don't get it why you always distort the truth in favor of AMD.

Anyway I hope Zen will be good especially in single-threaded so that we actually get a choice again.

 

[Abwx]

Sometimes you're so bent on showing the world + dog that they're wrong, you don't even notice people who are at least partially agreeing with you.

Possibly but then it s not like AMD related threads are not filled with people with dubbious agendas, sorry if i did misunderstand something..

On topic, we have people explaining us that Zen will be a small core, so according to them AMD will design a core that has half the throughput of an Excavator module, wich is completely ridiculous and at odd with any logicaly built argumentation, or just prove of clulessness.

As to put the discussion on track in respect of expectations people should check the Integer perfs of current chips before doing theories, integer IPC of a Haswell core is 41% better in Fritz and 28% better in 7 Zip than a Piledriver core, this put things in perspective about what they have to achieve to be competitive, that s not few but certainly not as much, and by far, as what some people think by there.

 

[Abwx]

Your own link shows that bulldozer loses to i7-860 in all single-threaded benches...and so does Piledriver. In fact they only heavy threaded benches and all of the are except Cine-Bench single-threaded. Of course the 920 loses more due to it's low clock. And I was usually following gaming benches from other sites where the i7-860 single-threaded performance was easily able to beat Piledriver in FPS. But that's actually off topic. I still don't get it why you always distort the truth in favor of AMD.

Anyway I hope Zen will be good especially in single-threaded so that we actually get a choice again.

Here an Integer single threaded test :

http://www.hardware.fr/focus/99/amd-fx-8370e-fx-8-coeurs-95-watts-test.html

 

[Vesku]

Even if Keller and team manage to work some magic I think getting around even IB IPC would be quite the turnaround for AMD. Even getting a combination of SB IPC and efficiency would be good especially if delivered on time.

 

[krumme]

We have a thread about zen. Fine. But both a72, Ares, a9, samsung and qcom custom cores is imo more interesting. Except ares they are probably all h2 products and driven by huge companies with financial and market punch. We should have separate thread about them too, and embrace the new players and arm tech. This discussion of amd tech is kind of from an old age not important any more imo. I would rather hear your impressions on the arm side and then let the other amd x86 rest until eg end h2. Sweepr done an excellent work here on recent a57 btw.

 

[mrmt]

At 2:40 you can find the "boil the ocean" part.
There's a specific question related to differences between A57 and K12 at 11:08, which is answered by Jim Keller in the presence of several AMD officials, including Lisa Su

I'll watch later, but as I said, unless you consider "high performance" something different from Intel top Xeon chip, there's a lot of miscommunication happening at AMD. AMD is telling their investors they are aiming for a top notch dense server product, not a conventional server chip, as Keller seems to be implying (by your statement).

 

[Enigmoid]

Keep in mind 4-5W per core is a different tier of performance compared to current ARM big core targets of ~1W per core (A57). So Keller's statements could easily be referring to an efficiency focused big core as we've been discussing rather than a king of the hill type of design. 4-5x W/core and ~4-5x the size would certainly be very powerful compared to A57.

A57 is not 1W per core when running without throttling.

Here an Integer single threaded test :

http://www.hardware.fr/focus/99/amd-fx-8370e-fx-8-coeurs-95-watts-test.html

You realize Fritz is junk right? Yet you still keep using it.

What is the calibration for Fritz?

P3 @ 1.0 Ghz = 480 points

Assuming linear scaling, the 2500 points that the 8350 gets at 4.0 Ghz would reduce down to 625 @ 1.0 Ghz.

Yay! 625/480 = 30% better IPC for Piledriver than Pentium 3.

Or perhaps the test is complete garbage. Its not an indication of performance, its not an indication of efficiency for any real world code.

 

[AtenRa]

Your own link shows that bulldozer loses to i7-860 in all single-threaded benches...and so does Piledriver. In fact they only heavy threaded benches and all of the are except Cine-Bench single-threaded. Of course the 920 loses more due to it's low clock.

http://anandtech.com/bench/product/434?vs=108
Have a better look, the only bench the FX8150 loses is in single tread Cinebench 10.
ps. dont bother with sysmark2007

And I was usually following gaming benches from other sites where the i7-860 single-threaded performance was easily able to beat Piledriver in FPS. But that's actually off topic. I still don't get it why you always distort the truth in favor of AMD.

Lets remember what you said and my response,

Wrong. i7-920 + higher models (Nehalem) and i7-860 + higher models (Lynnfield) existed before BD and are "8 Threaded" due to hyper-threading. In fact I still own such a Lynnfield and it still beats any FX in gaming...

Not on games with new Game-Engines like Frostbite 3, CryEngine 3+ etc etc. Also games using Mantle.

Ill give you a hint,

Thief is using the Unreal Engine 3, Intel CPUs are performing way better on this Game Engine and yet with Mantle the FX is faster both at default and OC than Nehalem. It is the same with every other Mantle game and any game using the FrostBite 2.5/3 and Cryengine 3+.

ps. Even at Bulldozer release date, many Games where equally performing at 1080p with Nehalem.

 

[Abwx]

A57 is not 1W per core when running without throttling.

You realize Fritz is junk right? Yet you still keep using it.

What is the calibration for Fritz?

P3 @ 1.0 Ghz = 480 points

Or perhaps the test is complete garbage. Its not an indication of performance, its not an indication of efficiency for any real world code.

Fritz is very convenient, and favour Intel over AMD, so what is your point.?.

We could use 7 zip but you would protest that it s anti Intel biaised...

As for calibration you re wasting your time with awfull clulessness since the score is in Kilo Nodes per second, the Pentium 3 score is just here for comparison with old uarches...

 

[Enigmoid]

Fritz is very convenient, and favour Intel over AMD, so what is your point.?.

Not very convenient since it uses ancient code that is not indicative of real world performance.

It doesn't favour intel over AMD. Intel ST integer performance is flat out faster in nearly every application which is mirrored in Fritz, without looking at relative percentages you can't say it favours intel.

We could use 7 zip but you would protest that it s anti Intel biaised...

7-zip and other decryption programs are very low IPC and latency sensitive. They are good comparisons for certain operations but terrible for others. In general 7-zip is one good benchmarking tool but others are required for a proper analysis of integer performance (ie steamroller posts gains over piledriver in pretty much everything but decryption).

As for calibration you re wasting your time with awfull clulessness since the score is in Kilo Nodes per second, the Pentium 3 score is just here for comparison with old uarches...

Both are in Knodes per second. P3 @ 1 GHz gets 480. BD @ 4.0 (probably a little higher with boost) gets ~2500. IPC is about 30% greater than P3.

This is not at all indicative of integer performance since modern integer code will run miles better on BD due to the inclusion of SSE2.

 

[Abwx]

Not very convenient since it uses ancient code that is not indicative of real world performance.

It doesn't favour intel over AMD. Intel ST integer performance is flat out faster in nearly every application which is mirrored in Fritz, without looking at relative percentages you can't say it favours intel.

Ok, so it s not usefull, but what is your own proposition..?.

7-zip and other decryption programs are very low IPC and latency sensitive. They are good comparisons for certain operations but terrible for others. In general 7-zip is one good benchmarking tool but others are required for a proper analysis of integer performance (ie steamroller posts gains over piledriver in pretty much everything but decryption).

You know that 7 zip show a much lower IPC differerence between AMD and Intel chips than Fritz , do you.??.

Indeed it is used by AT to check the server chips integer IPC.

Both are in Knodes per second. P3 @ 1 GHz gets 480. BD @ 4.0 (probably a little higher with boost) gets ~2500. IPC is about 30% greater than P3.

This is not at all indicative of integer performance since modern integer code will run miles better on BD due to the inclusion of SSE2.

We could use modern chess games that are ultra optimised up to SSE4 and even some Haswell exclusive instructions but you would still protest that they are even worse than Fritz when it comes to the results :

Stockfish compiled with GCC 4.8, with 3 versions , a classical one, a SSE4 optimized that get 3% with Intel and AMD and a version wich use Haswell s BMI instruction and that get a further 2%, the fastest version is used.

Houdini 4, there s two executable files, one for AMD and the other for Intel, the fastest version is used.

Houdini 4 is the soft that benefit the most from the upgrade to Haswell.

Theses are MT results but you can clearly see that AMD s chip behave better comparitively to Intel s in thoses modern softs, the ST perfs will be in adequation, get used to it, Integer perfs of AMD uarch is much stronger than its FP perfs comparatively to Intel s CPUs.