[Simbelmynë]

Hardware "theory" incoming

So I have been looking at the best price performance. 2 cores per memory channel seems to be optimal. A cluster of 8350k connected with 20gb infiniband (used on ebay) therefore seems to be in a good spot.

Since CFD benchmarks of this sort is difficult to find I would like some help in theorizing the actual performance.

Setup1:

7820X + motherboard + cpu cooler + PSU

Setup2:

2 x (8350K + motherboard + cpu cooler + PSU + 20 GB/s infiniband adapter)

I assume all other hardware such as RAM etc is the same for both setups.

The price difference is about 250 Euro in favor of Setup2, where I live.

As soon as we wish to scale it up further, the difference becomes larger since Setup1 then also needs the infiniband adapters etc.

For this comparison let's assume that it will not be up-scaled.

How would two 8350k fare against a single 7820X?

and

How would three 8350k fare against one 7820X? (This is approximately the ratio that we have, money-wise, the extra RAM may degrade the price ratio though)

Finally, the 7820X may be overclocked, but the 8350K may probably not (since we use a really cheap micro ATX motherboard).

[evcelica]

Simple Answer is:
A 7820k will be approx 2x as fast as a 8350 since it has 2x the memory bandwidth (4 memory channels vs 2), and memory bandwidth is mostly the limiting factor in CFD.
So: with that logic, three 8350's will be 1.5 faster than a single 7820K.

CPU overclock won't help too much, since you will be memory bandwidth limited, not CPU limited.

[Simbelmynë]

Hi evcelica,

Yeah, so this was the basis for my choice of the 8350k since it has a good ratio of price/bandwidth.

However, it is obviously not the only truth as can be seen in OpenFOAM benchmarks on various hardware
(In which I think the 7820X was overclocked, which helped it to beat the 7940X)

I think the question may be difficult to answer without someone with a 8350k, so perhaps your answer is the best one

1.5 times faster than the 7820X would be really nice for the same price though. Now, the question is if it can beat the 7301 in terms of price/performance. (Most likely not)

[flotus1]

Unless you are only looking at the maximum possible performance for a code that can run on distributed memory platforms with perfect scaling, there are more things to consider.

Other factors that come to mind In no particular order
1) single-core performance.
I3-8350k definitely takes the cake here. And mild overclocks are possible even with the cheapest Z370 boards.
2) shared memory capacity
not all workloads can run on distributed memory platforms. Having a single node with more memory and total performance can sometimes be beneficial
3) interconnects
Trying to scale out your setup, you will probably need a switch starting with 3 nodes. So adding another 7820x node will only cost you an additional IB card instead of a switch for the I3 setup.
4) overhead
the more nodes you have, the more work you have setting them up and keeping them running.

But getting back to the initial idea of this thread "So I have been looking at the best price performance"
Dual-socket Xeon E5 v2. CPUs are cheap, RAM is cheap, only motherboards are a bit more difficult to find. ASRock EP2C602 is still in stock in many shops and costs less than 300€ these days. With 2x8 cores and decent memory, this setup beats a 7820x in terms of parallel performance and is cheaper.