I think you're missing the point here - it's not about the gflops.
You're doing the overclock for more performance obviously but when stress testing you're trying to get it to fail so that's why you use the maximum amount of memory for bigger problem sizes. It's not about using specific memory sizes for optimum gflops. You're trying to stress test.
[email protected], 3.4 x 16 = 54.4GHz Theoretical maximum - what is this?
Sorry it is 54.4GFlops
Floating point operation (Flop) is the addition or multiplication of real numbers that can involve decimal fractions.
Modern processors involving SSE2 instructions in double-precision (64 bit) can perform 2 additions+2 multiplications per core in a cycle. So that is 4 Flops per core per cycle. So for a quad core that will be 16 Flops.
[email protected] means it is running 3.4billion cycles per second. Since quadcore cpu can perform 16Flops per cycle, we multiply that value with the no of cycles per second or cpu clock speed and in this case we end up with 16 x 3.4GHz = 54.4 billion Flops or 54.4GFlops or 54.4 billion Floating point operations per second.
Now 54.4 GFlops is the maximum output at 3.4Ghz.
However in real tests we can never achieve that value becaue of the hardware limitations such as ram latencies since IBT makes heavy use of ram.
As IBT involves solving system of linear simultaneous algebraic equations by means of Gaussian elimination method which I studied during my engineering degree; they involve multiplications and additions hence the Flops involved.
In IBT test we use RAM to store those equations which will be processed by cpu and result will be outputted in tems of GFLops. Sure using more ram means more equations will be stored and solved and hence the test will be longer and cpu will be under stress for greater duration.
However with low GFlops, cpu is carrying out less multiplications and additions on the equations per second. This means that cpu is performing at slower rate and hence the coretemps will also be lower. When the cpu is running slower, it's processing power is less hence in real world usage it will be able to pass IBT on lower vcore.
When cpu is performing closer to it's theoretical maximum, it is processing at greater speed. When it is doing so, it will also be releasing more heat. If your cpu cooler isn't good enough temps will hit the roof and you will fail the test. It will also require more vcore to stabilise.
So IBT can be looked at in 2 ways:
1) More ram places cpu under more loading time hence cpu is more stressed
2) Higher GFlops values means cpu processes/executes equations at faster rate and this will also be putting cpu under heavy stress.
Imho higher GFlops values beat higher amount of ram when stressing cpu.
A cpu can pass IBT test with large amount of ram at lower GFlops values but it will fail in a test with smaller amount of ram at higher GFlops values
. This is what I have found from my own experience using IBT
.
Here is the article on running LinPack by Perspolis (original author) which I have posted several times and is very good for understanding on how to run LinPack (Linx/IBT).
http://www.overclock.net/intel-cpus/645392-how-run-linpack-stress-test-linx.html
And here is the intel data sheet of GFlops for cpus at stock speed. If you carry out the calculation as mentioned above you will find that intel GFlops values agree
.
http://www.intel.com/support/processors/sb/cs-023143.htm#3
