Best CPU for neural networks: 4 fast cores vs 6 slow cores

Long story short:

for an intensive use of neural network toolbox in matlab what is best choice?

1. Intel Core i7 7700k, 4 cores @4.5ghz

2. Intel Core i7 6800k, 6 cores @3.6ghz

Thanks

[David J. Mack]

Just an idea: Also think about the graphics card as you could use it to speed up the computations (using gpuArray; at least for convolutional NNs this should be advantageous).

[Roberto]

Yes, I will consider it for sure for deep networks, but still I need to know the performance of neural network toolbox with those CPUs. Any info about that?

[Walter Roberson]

Memory path and cache would probably also be important.

Answer 1

[Walter Roberson]

Looking at

http://www.cpubenchmark.net/cpu.php?cpu=Intel+Core+i7-6800K+%40+3.40GHz&id=2785 6800 @ 3.4 GHz: 13569 CpuMarks

http://www.cpubenchmark.net/cpu.php?cpu=Intel+Core+i7-7700K+%40+4.20GHz&id=2874 7700 @ 4.2 GHz: 12277 CpuMarks

If we assume linear scaling with clock rate, then

6800 @ 3.6 GHz (3.6/3.4) * 13569 would be about 14367 CpuMarks
7700 @ 4.5 GHz (4.5/4.2) * 12277 would be about 13153 CpuMarks

so the 7700 system would be expected to be slower.

However, the 7700 CPU would be less expensive at current prices, and consume less energy. Also, the 7700 would be faster per core, by nearly a third, so for tasks that cannot be easily parallelized, it would be faster.

You would probably need to pay attention to memory speeds; the 3.6 GHz is quite a comfortable range these days, lots of support for that range, but the 4.5 GHz might require different memory chips to serve efficiently.

[Roberto]

Yes I've read tons of comparisons, for example:

Cinebench R15

singlecore

7700k --> 195 (+18%)

6800k --> 165 (+0%)

multicore

7700k --> 1022 (+0%)

6800k --> 1135 (+11%)

Geekbench

singlecore

7700k --> 5322 (+26%)

6800k --> 4224 (+0%)

multicore

7700k --> 21107 (+0%)

6800k --> 23013 (+9%)

It looks that the 6800k is much more outperformed by 7700k in singlecore tasks than 7700k is outperformed in multicore scenarios. Also considering that the 7700k is cheaper and a very good overclocker, it's probably be my first choice.

But, you know, these are only generic benchmarks... I was asking here hoping in some real world direct experience with neural network toolbox.

[Walter Roberson]

Alas, no-one ever sends me fun toys to test out :(

[Roberto]

Also tests with other cpus are very well accepted! I only need to know how neural networks behave in matlab with many cores available.

[Walter Roberson]

I only have one machine, a fall 2012 i7 MacBook with 4 cores. I could probably borrow a later iMac that I suspect is the same number of cores but probably a later i5 (I never thought to ask before about the CPU on it.)

You might want to consider using the Mathworks cloud service to benchmark a couple of different configurations. Or I am certain that Amazon has multiple configurations available that you could benchmark against.

[Roberto]

I was hoping to find info for free...

[Walter Roberson]

If you put together a representative benchmark that doesn't take too long and which spits out a couple of ratings of particular interest to you, then people might be willing to run it on a few different configurations.

[Roberto]

Not sure to understand. Do you mean here? You're suggesting to open a thread asking people to run some code and post the computation time with their configuration?

[Walter Roberson]

You could post it on this thread. You are getting a fair number of views, even if I have been the only one commenting so far.