Fudzilla: Bulldozer performance figures are in

[Despoiler]

it really seems to be broken...
people in Phoronix are very excited about trinity...
so...piledriver may be a fix?

http://www.phoronix.com/scan.php?page=article&item=amd_trinity_early&num=1

Someone mentioned AMD knows what the problem is and has fix. The fix won't be in the release BD though.

http://forums.anandtech.com/showpost.php?p=32385421&postcount=226

 

[Vesku]

This was in an Overclock.net thread:

http://imageshack.us/f/269/fxvsphiivsi5.png/

Hope this is true, shows an expected improvement from Thuban. We will find out in a day and a half I guess (assuming there are some midnight review releases).

 

[Despoiler]

Bulldozer shows up on 3dcenter benchmark listings

http://www.3dcenter.org/news/2011-09-22

 

[mosox]

How is this a fail I don't understand. What's wrong with all these benches?

http://www.cpubenchmark.net/cpu.php?cpu=AMD+FX-8150+Eight-Core

 

[Arkaign]

How is this a fail I don't understand. What's wrong with all these benches?

http://www.cpubenchmark.net/cpu.php?cpu=AMD+FX-8150+Eight-Core

Passmark is mostly bs.

Besides, highlight the thing, it says samples = "1". For all we know it was clocked at 6ghz under dry ice, who knows.

 

[-Slacker-]

Bulldozer shows up on 3dcenter benchmark listings

http://www.3dcenter.org/news/2011-09-22

Hmm ... one benchmark, the one from Hartware or whatever, shows the fx8120 at ~7.8 in the cinebench 11.5 bench (overclocked to 3.5 ghz)..

 

[Abwx]

As pointed by Dresdenboy , BD will need some compilers to optimize
for its architecture.

Otherwise , we ll keep on seeing things like this :

 

[NostaSeronx]

Passmark is mostly bs.

Besides, highlight the thing, it says samples = "1". For all we know it was clocked at 6ghz under dry ice, who knows.

Unless you can find out if Zambezi has a multiplier bug then the 1 sample is at stock settings(It shows measured speed at 3.616GHz)

 

[zlejedi]

How is this a fail I don't understand. What's wrong with all these benches?

http://www.cpubenchmark.net/cpu.php?cpu=AMD+FX-8150+Eight-Core

It has 1100T at level of i7 870 - I wouldn't treat such bench seriously

 

[NostaSeronx]

It has 1100T at level of i7 870 - I wouldn't treat such bench seriously

Passmark is an averaged out test
1100T isn't at the level of the i7 870

but one thing I would tell you is that
the i7 870 will run better in games and do havok better
while the 1100T will run better in multithread floating point and multimedia

i7 870 consumer
1100T aimed more at people who need more cores and know what programs actually yield benefits

1100T loses to the i7 870 in Integer Math(Games like Integer Math)
i7 870 loses to the 1100T in FP Calculations and SSE2 instructions and Encryption
1100T marginally is slower but very close to the 870 in Finding Prime Numbers, Compression, Physics, String Sorting(Games like Physics so 870 does better in both :hmm

If you can overclock the i7 870 you would be faster the 1100T @ same clocks in pretty much everything you do

i7 870 2.93GHz vs 1100T 3.3GHz
In most cases everything that is important on the i7 870 for the consumer who is buying it is faster than the 1100T
----
Hicooke hits 8GHz
LN2 Bulldozer on a Gigabyte:
http://valid.canardpc.com/show_oc.php?id=2040266

 

[3DVagabond]

As pointed by Dresdenboy , BD will need some compilers to optimize
for its architecture.

Otherwise , we ll keep on seeing things like this :

Very interesting! So, why are Opterons so much faster running Linux than Windows? Why is it just the opposite with Intel?

 

[-Slacker-]

Anyone gonna address this, or is this normal?

 

[Fox5]

Very interesting! So, why are Opterons so much faster running Linux than Windows? Why is it just the opposite with Intel?

I wonder what the compilation options and gcc versions used were...

But typically linux software makes heavier use of x87 floating point instructions, which AMD's more recent architectures have been faster at than Intel's architectures, so that might be why? Intel seems to focus on making the common instructions fast, while AMD was more 'well-rounded' in a sense. I don't think GCC has a good idea about the specifics of the architectures, so it compiles for X86 (or x86-64) in general, without having specific knowledge of stuff like "This cpu is super fast at int arithmetic and is faster to approximate a divide operation than to use its super slow divide operation."

Or perhaps Linux just makes better use of the opterons extra cores and snoop filter than windows does.

 

[-Slacker-]

^Doesn't superpi use x87 instruction sets?

 

[Olikan]

Or perhaps Linux just makes better use of the opterons extra cores and snoop filter than windows does.

the phenom arquitecture is just better for complex treads...
while core arquitecture is better for easy treads...

that's why, cray still didn't trade the brands

 

[lifeblood]

As already pointed , the current leaks show an amazingly low bandwidth for L1 Wrt/Cpy , two to three times slower than an X6 , assumed that the leaks are legit, so it remain to be seen if the current version is not literally unwillingly crippled..

JFAMD claimed BD has higher IPC than Phenom II. Yet everything we see says performance is worst. Assuming both JFAMD and the benchmarks are telling the truth, this would explain it.

So AMD might fix it with another spin of BD, or it might not get fixed until Piledriver. I bet AMD is moving heaven and earth to fix it in the server version of BD since that's where the big money is. Fine, but what about those who buy the borked workstation version of BD? It won't be me, that's for sure. This is Phenom all over again.

I really hope the leaked benchmarks are wrong and AMD has a truly competitive product. As it stands, BD is just not looking competitive to me.

 

[Idontcare]

JFAMD claimed BD has higher IPC than Phenom II. Yet everything we see says performance is worst. Assuming both JFAMD and the benchmarks are telling the truth, this would explain it.

So AMD might fix it with another spin of BD, or it might not get fixed until Piledriver. I bet AMD is moving heaven and earth to fix it in the server version of BD since that's where the big money is. Fine, but what about those who buy the borked workstation version of BD? It won't be me, that's for sure. This is Phenom all over again.

I really hope the leaked benchmarks are wrong and AMD has a truly competitive product. As it stands, BD is just not looking competitive to me.

If the benches are legit then BD does NOT have improved IPC over PhII...regardless the reasoning why...

There is a reason piledriver is not called bulldozer, if JFAMD wanted to talk about piledriver's IPC relative to Phenom II then he could have. But he did not.

Regardless the excuses to be conjured up...what is the relevance of, or value to come from, stating xyz, strongly and firmly, over and over again, if xyz has no basis in reality?

"Haswell will have 10X the IPC of Sandy Bridge!" (unless it doesn't, in which case, well maybe I was talking about Rockwell or something else entirely, who knows, guess you'll have to wait and see. K thx bye)

Its not like he went on record as saying "current steppings do NOT have better IPC, but we expect it to get there once we eliminate the critical bugs".

There is faceplant and then there is faceplant. Tomorrow will tell us which kind of faceplant we are looking at here.

Until then, bulldozer goes to 11.

 

[Ancalagon44]

Yeah I'm kinda excited to find out. In particular I'm looking forward to a thorough dissection of this performance issue, whatever it is, from Anand.

 

[lifeblood]

Regardless the excuses to be conjured up...what is the relevance of, or value to come from, stating xyz, strongly and firmly, over and over again, if xyz has no basis in reality?

JFAMD is in marketing and anything he says has a long line of conditions and legalese streaming behind it. In his mind the L1 reads and write may not be a part of IPC. Again, he/AMD can and will define IPC in a way that is most beneficial to them.

However, I accept your point. To you, me, and most others final overall performance is all that really matters.

Tomorrow we shall see indeed.

 

[Red Storm]

I just hope we get a competitive chip from AMD. Not even for my sake (no longer have a desktop), but for the sake of AMD making some money and helping to keep competition going in the CPU business.

 

[Ancalagon44]

JFAMD is in marketing and anything he says has a long line of conditions and legalese streaming behind it. In his mind the L1 reads and write may not be a part of IPC. Again, he/AMD can and will define IPC in a way that is most beneficial to them.

However, I accept your point. To you, me, and most others final overall performance is all that really matters.

Tomorrow we shall see indeed.

The performance of the processor is the sum of its parts, including cache.

So... if there is a problem with the L1 cache that affects performance and thus IPC, then something is wrong with the CPU itself. I dont see how JF could get around that.

As I've mentioned elsewhere, remember JFAMD is marketing. A) remember its his job to hype the product, B) remember he is not an engineer and thus does not actually know. If you asked a member of the engineering team at the time, you would likely have got a completely different answer.

 

[Vesku]

Put away the torches and pitchforks until you see Interlagos performance. JFAMD is a server guy, he's been very clear about not stepping into the enthusiast shark tank when it comes to claims.

 

[formulav8]

People Jf is only repeating what engineers told him. Period. Get over yourselves already. If the people who told him are wrong then what he said will be wrong as well. Ok?

 

[Abwx]

I wonder what the compilation options and gcc versions used were...

But typically linux software makes heavier use of x87 floating point instructions, which AMD's more recent architectures have been faster at than Intel's architectures, so that might be why? .

That has nothing to do with X87 being better on AMD ,
quite the opposite , in fact.

Chances are that blender for Windows is compiled with an intel compiler and that it doesnt use more than SSE1 and no SSE2 if the CPU is not a genuineIntel , reducing drastically the perfs.

Indeed , the K10 have powerfull FPU that was hardly matched
by Nehalem , unless the code is , well , unoptimised...

 

[garagisti]

Anyone gonna address this, or is this normal?

Read this comment on the following thread and move down further:
http://www.blendernation.com/2010/03/30/benchmark-test-results-using-blender-2-5-alpha-2/

Ivan Paulos Tomé

Hi dfelinto, and everyone at B.N.
1) Anandtech website change it's visual style and then some page reallocation ocurred during this transition. (3 days ago)
2) The graphics seems confuse at fisrt sight, sometimes we have to read full story articles to understand what's being done. Course, I will continue supporting them to do better tests using Blender as Benchmark tool.

The explanation of the variations of the Tests:
For Windows:
Windows Vista/7 uses a new kernel that is more preemptively scheduled than the old multitask kernel used in Win2000/XP/2003. Then the tasks are evenly distributed over the cores ( changing threads over time from processor to processor) . What's happened to Intel processors ?
The kernel changes threads over processors including HT ones that aren't real ones. My (later) suggestion to Johan was to disable HT in order to use only Phisycal processors, or set affinity manually, to improve test reliability.

In Linux:
Linux is a multitask kernel that does not change threads from processor to processor over time, and then, when a non Physical processor is found, the thread is tied to this processor until the task is finnaly done, but in this case, performance problems in HT (non Physical cores) may happen.

The test made, didn't take into account those differences, and it's not compiled to take advantage of some instructions. I've already sent an e-mail to Johan explaining those HT problems, processor scheduling problems and some more hints. That's the only way we can benefit from those tests before a decision to buy one or another Render Server. There is no Biased tests, the problem is that the setup needs to be carefully made to determine better results.

Ajusting Blender to fullfill a great number of processors is also another point, because some tiles are more demanding of processor power than others, and then, a Tile Preset for Threads seems impossible sometimes. AND, the eofw scene test is reaching its limit.

I'ts a good work to research more and more this Area. It's helpfull to persons that need to accomplish great Jobs with few Machines, Render Farmers, and even Gamers/Overclockers.

Good Vibes !