Bit-tech AMD six core review

[v0dka]

Why does the desktop keep coming into the discussion?

a) this is a server CPU
b) the test was done poorly (performance wise at least) I'm sure we can agree on that. Look at the data Hans de Vries posted for example.

[Donnie27]

Donnie,Hans was talking about server workloads where BW means a lot,not client one where it means very little. Give more BW to Istanbul and it will surely get better scores in a lot of workloads.How much better?Nobody knows until DDR3 server platform launches(my guess is quite good speed-up in spec rate tests at least,since these are mostly BW bound).
In the end,customers mostly look for perf/watt and AMD is doing very good at the moment. New SE 6core Opty at 2.8Ghz is already listed at HP's site,so the launch is imminent.With DDR3 things can only get faster for Istanbul,and with Magny Cours ,AMD will use simple yet effective tactic of core pair matching in direct connected MCM in order to get 2x more cores,2x more cache and almost 4 times more mem. BW than current Istanbuls(DDR3 vs DDR2 and 4 vs 2 channels).

Informal, I was talking about Servers, Conroe version of Xeon and Penryn version of Xeon, NOT limited to their Desktop brothers. More bandwidth would make this test look better. That will not translate to that much better performance in a lot of other apps. I agree no one knows how much better things would get. More bandwidth will help but IMHO, it will not change the outcome anytime soon. Sure I'm no expert.

I'll say it again, HPC folks aren't to worried about perf/watt and if that's what AMD is pinning their hopes on, they are in Deep DooDoo! There are already TV ads running for Intel/Dell/IBM/ and HP SERVERS guys. They're showing one blade rack replacing a whole room full of older servers. They've already planted the savings for Space, Power Draw, Temps and where are the AMD ads? This is one of the reasons AMD lost server market share.

Yes, I know Hans meant GB and not MB.

[generics_user]

i don't get it how people can defend this review?

it has so many obvious errors in component choice and offers no detailed information on the test setup...


you better take a look at the anandtech review published over a month ago:

http://it.anandtech.com/IT/showdoc.aspx?i=3571


Much bette insight in the hardware and benchmarks for the server market (who cares about GAMING performance on a server cpu )


bit-tech should stick to gaming hardware and try to improve their test setups instead of wasting their time on server hardware

[Particle]

Using avisynth? If so the bottleneck is there, before x264. People get 100% usage from 16 core with x264. There is something wrong with your process that causes this.

That must be it then. Yes, I use avisynth. Take note that I don't so much benchmark as look at the performance of actually useful tasks, and that means using avisynth to convert video into one of the like two formats the x264 encoder supports.

Example of AviSynth Script "File.avs":

DirectShowSource("Some File.avi")
ConvertToYV12()

Example of Encode Script "Encode.bat":

del "Temp Stats.log"

x264.exe --bitrate 7500 --pass 1 --stats "Temp Stats.log" --progress --threads 6 -o "Output.mkv" File.avs

x264.exe --bitrate 7500 --pass 2 --stats "Temp Stats.log" --progress --threads 6 -o "Output.mkv" File.avs

x264.exe --bitrate 7500 --pass 3 --stats "Temp Stats.log" --progress --threads 6 -o "Output.mkv" File.avs

[villa1n]

i don't get it how people can defend this review?

it has so many obvious errors in component choice and offers no detailed information on the test setup...


you better take a look at the anandtech review published over a month ago:

http://it.anandtech.com/IT/showdoc.aspx?i=3571


Much better insight in the hardware and benchmarks for the server market (who cares about GAMING performance on a server cpu )


bit-tech should stick to gaming hardware and try to improve their test setups instead of wasting their time on server hardware

qft

[informal]

Not to mention the MB/sec that should be GB/sec just as another obvious error...

It was a typo obviously .The man is writting a high level physics book btw,think 100 times before you point out "mistakes" such as this.

Since when was the X5570 3.33GHz?

Since X5570 does run all 4 cores with turbo(iirc).There are limits to it,but if the temps are below the specced limit ,all 4 cores run turboed.

i don't get it how people can defend this review?

it has so many obvious errors in component choice and offers no detailed information on the test setup...


you better take a look at the anandtech review published over a month ago:

http://it.anandtech.com/IT/showdoc.aspx?i=3571


Much bette insight in the hardware and benchmarks for the server market (who cares about GAMING performance on a server cpu )


bit-tech should stick to gaming hardware and try to improve their test setups instead of wasting their time on server hardware

Hit the nail on the head

[Hans de Vries]

Since when X5570 has became 3.33 GHZ? Last time I checked it was 2.93. Turbo Bust is very unlikely since all cores are active&busy in this bench.

The latter would be probably correct for you or me if we would buy a
single system, however,

The 3.33 GHz is the maximum frequency allowed to use for Spec 2006
when benchmarking an X5570 and there is nothing which forbids these
multinationals to actually use systems and CPU's which can and do run
at this frequency.

From a marketing standpoint they are more or less obligated to do some
cherry picking. If they wouldn't do so then their systems would look
inferior compared to those of their competitors.

Regards, Hans

[Shintai]

The latter would be probably correct for you or me if we would buy a
single system, however,

The 3.33 GHz is the maximum frequency allowed to use for Spec 2006
when benchmarking an X5570 and there is nothing which forbids these
multinationals to actually use systems and CPU's which can and do run
at this frequency.

From a marketing standpoint they are more or less obligated to do some
cherry picking. If they wouldn't do so then their systems would look
inferior compared to those of their competitors.

Regards, Hans

You do know how turbomode works right? Or are we now running spec numbers in dual or singlecore? Plus with turbomode you kinda calculate that the needed TDP headroom is there.

1-2 cores is 3.33Ghz in full turbomode on X5570. 3-4 cores is 3.2ghz. And with HT on and utilized its even harder to reach those. Unless its a "light" load.

[kl0012]

The 3.33 GHz is the maximum frequency allowed to use for Spec 2006
when benchmarking an X5570 and there is nothing which forbids these
multinationals to actually use systems and CPU's which can and do run
at this frequency.
Regards, Hans

Isn't max turbo restricted by some kind of internal logic which has no magic switch "on/off"? Untill now I didn't saw any evidence that Nehalem/Xeon can run 4 cores full loaded with max turbo on. Otherwise it makes no sence to mark this CPU as 2.93GHz.

[informal]

Isn't max turbo restricted by some kind of internal logic which has no magic switch "on/off"? Untill now I didn't saw any evidence that Nehalem/Xeon can run 4 cores full loaded with max turbo on. Otherwise it makes no sence to mark this CPU as 2.93GHz.

http://it.anandtech.com/IT/showdoc.aspx?i=3536&p=2

They say that x5570 usually works at higher clocks than def.,but 3.2GHz while achievable,is not seen that much.Also they stated the cores clock up to 3.06GHz usually.But if one would make the ambient and case temps lower,cores could theoretically clock up to 3.2Ghz as their table shows.

Xeon model Speed (GHz) Max. Turbo Max. Turbo(4 cores busy ) L3 Cache (MB) TDP (W)
X5570 2.93 3.33GHz 3.2GHz 8MB 95

[duploxxx]

Isn't max turbo restricted by some kind of internal logic which has no magic switch "on/off"? Untill now I didn't saw any evidence that Nehalem/Xeon can run 4 cores full loaded with max turbo on. Otherwise it makes no sence to mark this CPU as 2.93GHz.

the x5570 model can increase the multiplier by 2 with 3core/4core as long as it stays in TDP or increase multiplier by 3 for 1core/2core as long as it stays in TDP this is set in the CPU logic.

The Value(80W) series don't have turbo
The Volume(80W) series have +1 for 3c/4c and +2 for 1c/2c
The Perf(95W) series have +2 for 3c/4c and +3 for 1c/2c
The W (130W) series have +1 for 2c/3c/4c and +2 for 1c

so that is the reason why Hans stated upto 3.33GHZ on the x5570 since the x5570 falls into the 95W series
and Shintai, I am sure he knows....

[Chad Boga]

so that is the reason why Hans stated up to 3.33GHZ on the x5570 since the x5570 falls into the 95W series

Hans didn't state up to, he just stated 3.33Ghz as though it was a guaranteed constant.

I guess like so many members of AMDZone, he needs to over state what Turbo is bringing to the table.

[kl0012]

the x5570 model can increase the multiplier by 2 with 3core/4core as long as it stays in TDP or increase multiplier by 3 for 1core/2core as long as it stays in TDP this is set in the CPU logic.

The Value(80W) series don't have turbo
The Volume(80W) series have +1 for 3c/4c and +2 for 1c/2c
The Perf(95W) series have +2 for 3c/4c and +3 for 1c/2c
The W (130W) series have +1 for 2c/3c/4c and +2 for 1c

so that is the reason why Hans stated upto 3.33GHZ on the x5570 since the x5570 falls into the 95W series
and Shintai, I am sure he knows....

http://it.anandtech.com/IT/showdoc.aspx?i=3536&p=2

They say that x5570 usually works at higher clocks than def.,but 3.2GHz while achievable,is not seen that much.Also they stated the cores clock up to 3.06GHz usually.But if one would make the ambient and case temps lower,cores could theoretically clock up to 3.2Ghz as their table shows.

Xeon model Speed (GHz) Max. Turbo Max. Turbo(4 cores busy ) L3 Cache (MB) TDP (W)
X5570 2.93 3.33GHz 3.2GHz 8MB 95

So guys, you both support my argument that x5570 can't run 3.33GHz with 4 cores busy. So how Hans can be right with his statement?

[SEA]

Hm, there is one thing that would allow to run 4 thread task at 3.33. That would depend on how windows would dispatch loads, e.g. this way: 2 real cores + 2 virtual ones.
So in case of windows it would run almost half of time @ 3.33 and rest at 3.2Ghz (if TDP allows, of course)

And apparently bit-tech used mostly 4 or less threaded tasks.

[Shintai]

Hm, there is one thing that would allow to run 4 thread task at 3.33. That would depend on how windows would dispatch loads, e.g. this way: 2 real cores + 2 virtual ones.
So in case of windows it would run almost half of time @ 3.33 and rest at 3.2Ghz (if TDP allows, of course)

And apparently bit-tech used mostly 4 or less threaded tasks.

Windows loads the cores correctly. The extra logical cores per real core is loaded last.

[informal]

Well the max turbo of 3.33Ghz is achievable on less than 4 cores and I believe there are tests that don't fully load all 4 cores in spec rate suit.So yes,the CPU could run two active cores at 3.33GHz if the load on the cores allowed it. The top scores we see posted are usually tuned to the max(ie. used low ambient and case temps,max. tuned via compiler switches etc.).It's legit of course,people use everything they can without breaking the rules in order to take the top spot.

[SEA]

Windows loads the cores correctly. The extra logical cores per real core is loaded last.

They used Windows Vista Ultimate x64 SP1
I'm not sure if it is correct for non-server versions of windows.

Anyway, 3.2 or 3.33 Ghz - not a big difference vs 2.6 Ghz

[Shintai]

They used Windows Vista Ultimate x64 SP1
I'm not sure if it is correct for non-server versions of windows.

Anyway, 3.2 or 3.33 Ghz - not a big difference vs 2.6 Ghz

Vista and Windows Server 2008 is the same kernel and codebase.

[Lightman]

I wonder if they bothered to change number of render threads in LightWave 3D.
The software is auto-configuring that option depending on number of cores at the installation (or during startup), then writes it to cfg file. It only allows 1,2,4,8,16 threads, so it's highly possible that 12 core machine was still rendering using 8 threads! Manually adjusting that value to 16 threads would surely improved Istanbul score.

Still 16 render threads are not enough for 12 real cores! I'm basing that statement on personal experience with various 3D scenes on my Quad CPU. For my Quad best setting depending on scene was 8 or 16 threads! Almost always 4 threads was slower than 8! Reason being that render screen is split in equal parts (horizontal bands) according to number of render threads. Given that your scene is consisting of blue sky up to half of the picture and massively complex city/forest for the other part, then with 4 threads rendering top part utilizing 2 cores would finish way ahead of bottom part. There is no dynamic thread utilization as seen in Cinebench!
Granted, my findings were on LW 8.5, so NewTek could change the way renderer works, but they still kept 16 threads max. count in version 9.6 so I doubt it!

[SEA]

Vista and Windows Server 2008 is the same kernel and codebase.

Sure, like windows XP and Windows 2003 server. Which do dispatch DIFFERENTLY....

btw I'd like to learn how vista and other OS do scheduling... Any link to this? I could not find, particularly could not find anything on how windows makes difference on virtual and real cores task scheduling. Can you support your words with some useful link to MS?

[Shintai]

Sure, like windows XP and Windows 2003 server. Which do dispatch DIFFERENTLY....

btw I'd like to learn how vista and other OS do scheduling... Any link to this? I could not find, particularly could not find anything on how windows makes difference on virtual and real cores task scheduling. Can you support your words with some useful link to MS?

XP and 2003 is only the same for x64. x86 is different.

Its no diffrent than when you do SMP and Multicore scheduling. Today its mostly done via ACPI to tell the OS who is who. If you load a HT system. i7, Xeons etc. You will also see its 0, 2, 4 and 6 that gets loaded first. Windows 7/2008 R2 should further improve this. XP was the first Windows OS to be Hyperthreading aware. Or rather aware of the differences on logical cores.

Ofcourse you can always mess it up by hardcoding affirmity in your app.

[SEA]

Shintai
OK, I see.
So every even core is counted as real core. And it assumes that BIOS enumerates them it this order. And if another BIOS would do differently? Or ACPI gets information on how they enumerated before starting scheduling?

XP was the first Windows OS to be Hyperthreading aware.
And since SP2 it is NUMA aware.

Still i don't see how all this would garantee 3.2 rather 3.33 Ghz...

[Hans de Vries]

If a Spec_Int_rate or a Spec_FP_rate submission specifies that Turbo mode
is used with speeds up to 3.33 GHz then this is what one (conservatively) has
to assume. The submitter has the option to attach notes which specify any
specific deviations.

It's not the task of the reader to go into any unspecified and uncontrollable
details here. The submitter is allowed to use any cherry picking, hardware-
settings, bios or kernel programming tricks to get the frequencies as close to
the maximum of 3.33 GHz as he can.


Regards, Hans

[Shintai]

Shintai
OK, I see.
So every even core is counted as real core. And it assumes that BIOS enumerates them it this order. And if another BIOS would do differently? Or ACPI gets information on how they enumerated before starting scheduling?

Thats how it worked long ago in XP32. In 2003 you got an API called GetLogicalProcessorInformation.

GetLogicalProcessorInformation returns data that describes:
The physical-to-logical relationship for each physical processor in the system.
The logical processor mask for each NUMA node in the machine, if the system has a NUMA architecture.

So now even your application and see if its a HT CPU or not. So if coded correctly it would simply make sure that the application only ran on 1 instance per physical core if HT results was proven to be bad.

[kl0012]

Well the max turbo of 3.33Ghz is achievable on less than 4 cores and I believe there are tests that don't fully load all 4 cores in spec rate suit.So yes,the CPU could run two active cores at 3.33GHz if the load on the cores allowed it. The top scores we see posted are usually tuned to the max(ie. used low ambient and case temps,max. tuned via compiler switches etc.).It's legit of course,people use everything they can without breaking the rules in order to take the top spot.

Hardly belivable considering 8 instances of app runnig on 4 cores (+HT). I rather belive that most of the time it ran at 2.93 GHz with high CPU utilization.

If a Spec_Int_rate or a Spec_FP_rate submission specifies that Turbo mode
is used with speeds up to 3.33 GHz then this is what one (conservatively) has
to assume. The submitter has the option to attach notes which specify any
specific deviations.

It states that turbo mode was active. It doesn't mean that CPU was in turbo mode during the entire test.

It's not the task of the reader to go into any unspecified and uncontrollable
details here. The submitter is allowed to use any cherry picking, hardware-
settings, bios or kernel programming tricks to get the frequencies as close to
the maximum of 3.33 GHz as he can.


Regards, Hans

Care you to explain how it is possible to ajust bios settings/kernel in such way, that additional 133-266MHz of frequency will be equivalent to one cpu core (since we already know that 3.33GHz isn't possible when app loads all 4 cores of x5570).