Linpack Challenge

[Idontcare]

@Aigo: I thought Magny Cours will have 12 physical cores w/ quad-channel DDR3? 2S Magny Cours will have 24 cores total. 2S Gainestown platform has 8 physical cores and 16 logical cores. I think you were thinking of dual Istanbul (12 cores) which indeed lose flat out to Nehalem-EP (8 physical cores / 16 logical cores). Nehalem-EX will be a completely different beast, of course.

I could be wrong on the Magny Cours.. I haven't really been a big fan of many cores.

P.S. There is a timely article by Johan.

http://it.anandtech.com/IT/showdoc.aspx?i=3681

Yes, and there is a guy over on XS, memory is failing me at the moment, who has a Magny-Cours ES 2S rig and can overclock the chips to 3GHz on silly little voltage.

He ran some wprime about a month ago and posted it, maybe older than that even.

So yeah, I got to imagine his 24 core rig would do "well" in this app.

 

[JFAMD]

@Aigo: I thought Magny Cours will have 12 physical cores w/ quad-channel DDR3? 2S Magny Cours will have 24 cores total. 2S Gainestown platform has 8 physical cores and 16 logical cores. I think you were thinking of dual Istanbul (12 cores) which indeed lose flat out to Nehalem-EP (8 physical cores / 16 logical cores). Nehalem-EX will be a completely different beast, of course.

I could be wrong on the Magny Cours.. I haven't really been a big fan of many cores.

No, you are correct. I think people are making up numbers here to state a position.

 

[aigomorla]

Wait the magny Cours was the AMD Hexcore?

Am i getting the processors mixed up....

If i am then i appologize... i have no idea how it will do.

But i did see a dual hexcore system from AMD @ 2.X ish and it wasnt very impressive on points. :\

 

[Shmee]

I ran 20 passes for the screenshot for evga forums to show a stable 4+ GHz oc, but only got 50 GFlops cuz HT I turned on. currently at 196x21.

 

[zsdersw]

I think people are making up numbers here to state a position.

Color me "not surprised" that you'd think as such.

 

[Idontcare]

Wait the magny Cours was the AMD Hexcore?

Am i getting the processors mixed up....

If i am then i appologize... i have no idea how it will do.

But i did see a dual hexcore system from AMD @ 2.X ish and it wasnt very impressive on points. :\

Aigo, Magny-Cours is 12 cores in a single-socket created by MCM'ed Hexcore IC's.

So a 2S rig would give you 24 cores. Clockspeeds will be low of course since 12cores * X Watts/core < 150W/socket so you have about 12.5 W/core of thermal budget (max) per core to hit whatever clockspeeds you want.

 

[JFAMD]

Color me "not surprised" that you'd think as such.

It happens on both sides.

 

[zsdersw]

It happens on both sides.

That doesn't justify it.

 

[JFAMD]

I am not justifying it, just stating that it happens. I don't have the benefit of anonymity so I have to state what I know. Others have the luxury of gravitating to any statement that they have seen (from any source) and lock on that as fact. As humans it is our nature to grasp onto statements that we want to believe and ignore those that conflict with our beliefs. It's just the way we are wired.

 

[BababooeyHTJ]

Case in point why F@H isnt a good stability indicator then. You probably still give them alot of garbage WUs back with that rig...

This Linpack really makes my box steam like no other tool. I'll probably do a full run myself later this weekend.

silent data corruption

I'm not a fan of Linpack. It pulls more of a load than your processor will ever see and while doing that cause more vdroop. If that lower vcore that you will never see again causes you to fail are you truly unstable? I noticed that even the second poster here failed.

Now that said. I have passed Linpack and still had stability issues that I found with Prime. With my old Wolfdale I could be 40+ run Linpack stable and still fail Prime small fft after about an hour. I think adjusting my GTLs fixed that. Secondly with my Yorkfield after I upgraded to 8GB of ram I was once again 100&#37; Linpack stable but my rig was acting flaky during HCI Memtest but still didn't show errors and failed a custom Prime run in under 2 (yes two) minutes.

Like I said I'm just not a fan of Linpack.

 

[n7]

I'm not a fan of Linpack. It pulls more of a load than your processor will ever see and while doing that cause more vdroop. If that lower vcore that you will never see again causes you to fail are you truly unstable? I noticed that even the second poster here failed.

Now that said. I have passed Linpack and still had stability issues that I found with Prime. With my old Wolfdale I could be 40+ run Linpack stable and still fail Prime small fft after about an hour. I think adjusting my GTLs fixed that. Secondly with my Yorkfield after I upgraded to 8GB of ram I was once again 100% Linpack stable but my rig was acting flaky during HCI Memtest but still didn't show errors and failed a custom Prime run in under 2 (yes two) minutes.

Like I said I'm just not a fan of Linpack.

As myself & others have said before, you should always be able to pass Linpack just fine @ stock speeds.

The fact that you cannot always @ OCed speeds basically points to instability, or at the very minimum, not the same level of stability as stock.

You can argue all you want, but if you can pass it @ stock speeds yet not @ OCed, i find it rather incredulous that it's somehow dismissed as "being too stressful, etc.".

Also, it's not the only stability test you should run.
Different tests show different issues.

LinX/IBT do a good job with testing stability of CPU, more than pretty much anything else.
Also i'd say likely second best for checking RAM.

P95 is better for isolating issues though.
Small FFTs for CPU
Large FFTs can stress RAM, but it's rather unreliable for larger amounts of memory.

I found that with quads on my P45, Large FFTs was excellent for finding issues with the motherboard's stability, particularly VTT or NB issues, & yes, it found issues LinX did not (just like LinX finds CPU issues P95 does not).

HCI Memtest (ideally as many instances as logical cores) is easily the best i've seen for RAM, especially with larger amounts of RAM.

Memtest86+ is good for memory errors, or basic RAM OC checking (test 5).
Not so good for checking for stability though.

I guess what i'm trying to say that there are many stresstesting programs that have value in different ways.

You can't simply dismiss one because of one thing though...they all have advantages; i generally run at least LinX/P95, & when working on RAM (or NB w/ P35/P965 in my cases), HCI.

 

[BababooeyHTJ]

As myself & others have said before, you should always be able to pass Linpack just fine @ stock speeds.

The fact that you cannot always @ OCed speeds basically points to instability, or at the very minimum, not the same level of stability as stock.

You can argue all you want, but if you can pass it @ stock speeds yet not @ OCed, i find it rather incredulous that it's somehow dismissed as "being too stressful, etc.".

Also, it's not the only stability test you should run.
Different tests show different issues.

LinX/IBT do a good job with testing stability of CPU, more than pretty much anything else.
Also i'd say likely second best for checking RAM.

P95 is better for isolating issues though.
Small FFTs for CPU
Large FFTs can stress RAM, but it's rather unreliable for larger amounts of memory.

I found that with quads on my P45, Large FFTs was excellent for finding issues with the motherboard's stability, particularly VTT or NB issues, & yes, it found issues LinX did not (just like LinX finds CPU issues P95 does not).

HCI Memtest (ideally as many instances as logical cores) is easily the best i've seen for RAM, especially with larger amounts of RAM.

Memtest86+ is good for memory errors, or basic RAM OC checking (test 5).
Not so good for checking for stability though.

I guess what i'm trying to say that there are many stresstesting programs that have value in different ways.

You can't simply dismiss one because of one thing though...they all have advantages; i generally run at least LinX/P95, & when working on RAM (or NB w/ P35/P965 in my cases), HCI.

Eh, you are right. I allways run Linpack but not the 40 or so runs that I used to since a long run of Prime small and Large FFT is usually a good indication that I'll pass Linpack. It is a good quick stress test but like you said I have just never found it very good for isolating errors.

 

[zsdersw]

I am not justifying it, just stating that it happens. I don't have the benefit of anonymity so I have to state what I know. Others have the luxury of gravitating to any statement that they have seen (from any source) and lock on that as fact. As humans it is our nature to grasp onto statements that we want to believe and ignore those that conflict with our beliefs. It's just the way we are wired.

Then you state the obvious, which isn't terribly useful.

The skill of critical thinking is lost on many people, including many here.

 

[Idontcare]

Then you state the obvious, which isn't terribly useful.

The skill of critical thinking is lost on many people, including many here.

Careful there, you run risk of stating the obvious as well.

And in looking over your posts here I can't readily discern whether you've really added anything terribly useful to the thread beyond badgering and baiting a fellow poster.

 

[zsdersw]

And in looking over your posts here I can't readily discern whether you've really added anything terribly useful to the thread beyond badgering and baiting a fellow poster.

Sorry, I'm out of caring minutes.

 

[Idontcare]

This LinX benchmark is an intriguing case-study for extracting thread-scaling details for a given architecture and platform. The application appears to be nearly 98&#37; parallelized and it is just fine-grained enough that memory bandwidth is a factor.

Would be cool if one of you nehalem and PhII owners could provide your GFlops per thread (run app with 1 thread, then 2, then 3, etc until you max out your threads for your cores) numbers, I'll crunch the data and report back if you give me the data.

 

[Hyperlite]

This LinX benchmark is an intriguing case-study for extracting thread-scaling details for a given architecture and platform. The application appears to be nearly 98% parallelized and it is just fine-grained enough that memory bandwidth is a factor.

Would be cool if one of you nehalem and PhII owners could provide your GFlops per thread (run app with 1 thread, then 2, then 3, etc until you max out your threads for your cores) numbers, I'll crunch the data and report back if you give me the data.

Neat. I'll have RAM back tomorrow and run it on my AII 620.

 

[Idontcare]

Neat. I'll have RAM back tomorrow and run it on my AII 620.

Cool, in case you didn't see the option, to change the number of threads LinX spawns just go into the Settings menu option and change the number circled in red in this screenshot:

 

[Hyperlite]

Cool, in case you didn't see the option, to change the number of threads LinX spawns just go into the Settings menu option and change the number circled in red in this screenshot:

10-4. thanks.

 

[lopri]

Cool, in case you didn't see the option, to change the number of threads LinX spawns just go into the Settings menu option and change the number circled in red in this screenshot:

I may try that. How much would L3 cache and/or the amount of memory matter? I know that Linpack will attempt to maximize the available cache and memory (which in itself runs into diminishing returns), so should I halve the amount of RAM while running with 2 cores instead of 4? Obviously I don't have a control of amount of L3, so that's another question mark.

 

[Idontcare]

lopri I left the ram available the same, simply varied the number of threads.

Ideally if we really wanted to extract the maximum amount of precision regarding the intricacies of the microarchitecture we would restrict the footprint of the application to an equivalent "share" of the shared resources (L3$, dram) as we scaled the number of threads...but this particular application appears to change the amount of work being done if you change the amount of ram it has access to and we actually want to hold that constant across our tests.

So the lesser of the two evils here is to keep the same amount of ram/cache across all tests and merely vary the number of threads spawned.

 

[Hyperlite]

lopri I left the ram available the same, simply varied the number of threads.

Ideally if we really wanted to extract the maximum amount of precision regarding the intricacies of the microarchitecture we would restrict the footprint of the application to an equivalent "share" of the shared resources (L3$, dram) as we scaled the number of threads...but this particular application appears to change the amount of work being done if you change the amount of ram it has access to and we actually want to hold that constant across our tests.

So the lesser of the two evils here is to keep the same amount of ram/cache across all tests and merely vary the number of threads spawned.

so run tests at...1024? 2048?

 

[Idontcare]

I did max ram just as lopri spelled out in his request at the top.

Only difference from his procedure is that I varied the thread number.

Oh and I also enable the "Auto-save log" checkbox in the settings window because I wanted to average the GFlops value from all the runs versus just using the absolute peak value. That part doesn't really matter though, purely second-order for our purposes.

(if you do elect to record all the runs for purposes of averaging then you can reduce the number of tests from 20 to 10 as 10 runs is plenty of data for the purpose of averaging and a single-thread test will take a long time)

 

[lopri]

I wasted some time being stupid. I varied the number of cores using Task Manager's affinity select without thinking of LinX's built-in thread selecting, and for some reason the results are all over the place. For example I assigned core 0/1/2 to LinX via Task manager and started it, then it returned results all over the place. (33 GFlops the first run, 26 GFlops the second run,.. etc.) I don't know whether that's a 'normal' behavior but the results were certainly garbage.

I guess I will try again using the LinX, or changing Window's boot config. Unfortunately my board's BIOS doesn't let me choose the number of cores which would have been my first choice.

 

[dank69]

Only did 5 runs each...

i7920(D0)@4.03 GHz

Compiled google doc for IDC:
LinXi920D0

SS or it didn't happen:
1 thread (HT)


2 threads (HT)


3 threads (HT)


4 threads (HT)


5 threads (HT)


6 threads (HT)


7 threads (HT)


8 threads (HT)


4 threads (No HT)


Prime95 39hr+ stable: