IBT

[WingZero30]

IBT is a front-end to the Intel Linpack benchmark http://software.intel.com/en-us/articles/intel-math-kernel-library-linpack-download/. The included matrix operations may very well include trigonometric functions in addition to multiplications and additions, I have not seen an exact specification of the Linpack subroutines included in the benchmark.

It's solving simultaneous eqns according to http://en.wikipedia.org/wiki/Linpack as I described above. If this isn't the case I'd like to know.

It seems like it is though, as e.g. in LinX you can choose the number of eqns explicitly.

Choosing more ram will result in more simultaneous equations which will be converted into matrix form, i.e the matrix size will be bigger.
Once the matrix is stored in ram, IBT then makes use of Gaussian Elimination method in order to solve the matrix, hence Flops being carried out by the CPU.

When I did Gaussian elimination back in uni, there were no trignometric steps involved. However division and subtraction were to some extent carried out if I recall correctly.

So I think IBT will involve division and subtraction subroutines but they won't get as much used up as multi-add.

Simultaneous equations can be solved by multiplying the equations by a same number to make an unknown say x have same value. Then the two x can be added if one is +ve and the other is -ve hence taking it out of the equations thus simplifying equations to solve them.

Multi-add is more popular than say division or trignometric functions because the former is quicker to implement per cycle whereas the latter could/will take more cycles hence the process will be slower. This means that cpu will be slower hence it won't be as much stressed as when using multi-add.

Even when they talk about GFlops of PS3 or Xbox360, they are usually referring to multi-add instructions

 

[WingZero30]

Thanks, so it does not seem to include any trigonometric functions. This is a major difference from prime95, which needs trigonometric functions for the Fourier transformations. It would be good to have a prime95 version that used AVX---this could be the ultimate test.

Yes you are correct in saying prime95 makes use of fourier transforms which is based on fourier series and include trignometric functions plus calculus. From mathematical standpoint, fast fourier transforms or FFTs is more complex to solve than gaussian elimination matrix. But because FFTs involve trignometry and calculus, cpu takes longer to calculate them than simple matrix method.

IBT is also cyclic in nature where it places tremendous load on cpu and when it outputs result, the load is temporarily lightened and then it goes back up.

Prime95 on the other hand is more about constant loading where it is solving fast fourier transforms.

I think it is the way Prime95 and IBT are programmed to run determine which is more stressful on cpu. IBT for cpu overclock whereas Prime95 is more for long term overall system stability and I would say neither should be neglected.

It is just that Prime95 and IBT are two very different stress tests and can't be compared directly.

 

[Plec]

>snip<

Succinct and informative WingZero.

I'm going to quote you in my above post as you’ve added the substance to my vagaries…

 

[betaspin]

It is just that Prime95 and IBT are two very different stress tests and can't be compared directly.

Still, the levels of stress achieved with Prime95 (small FFTs) and IBT/LinX without AVX appear to be very similar, at least so far as the CPU temperatures are concerned. Small FFTs can be vectorised very efficiently, so if Prime95 starts using AVX (i.e., vector-oriented FMAs), the stress will increase considerably.

I have to admit that my favourite test is the OCCT PSU test (they seem to have included linpack with AVX in the latest beta), one can achieve much the same by running IBT and Furmark at the same time. 1h of this appears to be a pretty good indication of stability and cooling capability.

 

[WingZero30]

WZ30 can we make a IBT sticky guide please? I'm sick of seeing confusion over this truly excellent stability prog.

I think we need to ask Mods to make this thread a sticky so anything IBT related can be discussed here so we can all gain better understanding of this stressful program.

I would also like to have a Prime95 sticky aswell. I think if we are using these two famous stress tests, then we should atleast have some understanding on how to optimise the settings in order to run them more accurately and efficiently on our system.

 

[WingZero30]

Still, the levels of stress achieved with Prime95 (small FFTs) and IBT/LinX without AVX appear to be very similar, at least so far as the CPU temperatures are concerned. Small FFTs can be vectorised very efficiently, so if Prime95 starts using AVX (i.e., vector-oriented FMAs), the stress will increase considerably.

If the programming of Prime95 can be modified to include AVX as you stated, then it should increase the stress on cpu.

Even I personally feel that since Prime95 is already solving a very complex mathematical problem involving FFTs, if only it can be modified to make it more stressful, then that will be a bonus.

Perhaps you can email this suggestion to creaters of prime95 and see what they have to say.

 

[paradisiac]

well im now on run 26/50 @ 4.8GHz without problems.
131.***secs 124.*** Gflops.

 

[betaspin]

Perhaps you can email this suggestion to creaters of prime95 and see what they have to say.

This has been discussed on the mersenne forum, see http://www.mersenneforum.org/showthread.php?t=14335. There will be an AVX version, but it may take some time to appear.

 

[paradisiac]

completed 50 runs of IBT @ 4.8GHz (1.39v) 131.***secs 124.*** Gflops Max temp 87c (but averaged 77c) which is lower temps than i was getting at 4.4GHz with my Freezer 7 Pro.
will slowly lower the vcore now, just whacked it to 1.39v to see if it would be enough and seems to be so now is the tweaking stage. will run it as is for now as i believe im still ok for 24/7 (although it only get a max 4 hours a day usage) vcore? see if i get any issues, if not then will attempt to lower it over the weekend.

 

[j.col]

completed 50 runs of IBT @ 4.8GHz (1.39v) 131.***secs 124.*** Gflops Max temp 87c (but averaged 77c) which is lower temps than i was getting at 4.4GHz with my Freezer 7 Pro.
will slowly lower the vcore now, just whacked it to 1.39v to see if it would be enough and seems to be so now is the tweaking stage. will run it as is for now as i believe im still ok for 24/7 (although it only get a max 4 hours a day usage) vcore? see if i get any issues, if not then will attempt to lower it over the weekend.

well done mate, 50 runs and you are well stable.
Tbh i just do 25 runs and i have never had a bsod, but its personal preference.

 

[joeyjojo]

Yeh paradisiac that's sweet, 4.8! People are saying 1.4 V is OK, time will tell I suppose.

These chips really are immense. Your chip is doing 6 times the number of operations per second as my E7200

 

[paradisiac]

well done mate, 50 runs and you are well stable.
Tbh i just do 25 runs and i have never had a bsod, but its personal preference.

i did 50 on my 4.6GHz (before using correct settings) and it failed on run 48 so i will stick with 50 just incase, took around 2 hours so was a pretty good test and a lot quicker than prime.

Yeh paradisiac that's sweet, 4.8! People are saying 1.4 V is OK, time will tell I suppose.

These chips really are immense. Your chip is doing 6 times the number of operations per second as my E7200

yeh can probs push for 5GHz @ 1.4-1.41v in bios, which will be 1.38-1.41v in windows. maybe even achieve it on less, but 4.8GHz is a pretty nice 'FREE' increase.

 

[joeyjojo]

Well it's not really free. Using the trusty OC calc at http://extreme.outervision.com/tools.jsp

3300, 1.21 = 95 W = £108/year
4800, 1.39 = 182 W = £207/year

at 13p/kWh, £100 a year more at full load 24/7.

 

[paradisiac]

Well it's not really free. Using the trusty OC calc at http://extreme.outervision.com/tools.jsp

3300, 1.21 = 95 W = £108/year
4800, 1.39 = 182 W = £207/year

at 13p/kWh, £100 a year more at full load 24/7.

hmm........i see your point. but as my computer is on at max 4hours a day, and thats not every day, then the increase is probably here nor there.

although, going by that logic too, if i stuck on stock i wouldnt of need £47.99 worth of cooler either

 

[WingZero30]

This has been discussed on the mersenne forum, see http://www.mersenneforum.org/showthread.php?t=14335. There will be an AVX version, but it may take some time to appear.

I think the Prime AVX version will be similar to IBT AVX supported version in that sandybridge, bulldozer, ivybridge , future cpus etc will be able to utilise it and not older core 2 like my Q6600 which only utilise instructions upto SSSE3.

So even if I use the Prime AVX on my Q6600, it won't make a difference compared to the current Prime95 in terms of stressing my cpu any further.

Edit: Here is a further thread on the Prime95 AVX

http://www.mersenneforum.org/showthread.php?t=15732

It seems some of the FFTs sizes have been coded to use AVX but still seems to be in development phase.

Just got numbers for the 2K FFT:

SSE2: 100 iters in 1.110 ms
AVX: 100 iters in 0.682 ms

I'm not sure if these numbers will extrapolate to large FFTs. These small FFTs operate completely in the L1 cache.

AVX again seems to be almost twice as fast as SSE2. Kind a like doubling Flops from 4 to 8 in double precision.

With cpu solving the FFTs faster, that sure is bound to make the cpu much hotter

 

[joeyjojo]

hmm........i see your point. but as my computer is on at max 4hours a day, and thats not every day, then the increase is probably here nor there.

although, going by that logic too, if i stuck on stock i wouldnt of need £47.99 worth of cooler either

I thought that would be the case. I just love doing calculations

(Need to get a job)

 

[paradisiac]

so what exactly is a Gflop, and does it actually do anything in normal pc usage? as in, does having higher Gflops mean it can do stuff quicker?

 

[joeyjojo]

http://en.wikipedia.org/wiki/FLOPS

FLOPS is a abbreviation for Floating Point Operations per Second. CPUs perform either integer or floating point operations, the former generally being less complex and thus faster. I'm not sure where instruction sets come in, maybe somebody else knows.

As I said above, "one hundred and nineteen billion floating point operations for second is pretty damn incredible".

More FLOPS means you can do stuff quicker. Very few tasks will reach the peak values though, including Prime 95. This is where IBT really is exceptional, as it can load the CPU to 80%+ of it's theoretical maximum.

 

[paradisiac]

so in real world terms, it makes no difference, but in stability testing it is one of the more complx tasks so should show issues quicker?

 

[joeyjojo]

The GFLOPS column in the window shows you how many operations were performed per second on average (in billions) while stress testing. As you use more memory, the program gets better at utilising your CPU, so the number increases. It is a measure of how fast the CPU performed the stress test, nothing more and nothing less. The peak performance of the chip is determined by the clock speed only. Which is why you're overclocking in the first place right?

The CPU is always going to be faster in the real world as the clock speed increases

IBT is fast and hot because it utilises the CPU so fully.