Intel CPUs performance

Most people recommends to disable hyper-threading for tasks such as signal processing on Intel CPUs.

So,there will be no difference between two CPUs, if they have same core architecture and same number of actual cores.

Not only frequency matter: instructions set (AVX2 for example) and the core internal architecture are important.

Many people have issues related to thermal capacity of the box: brick like format may decrease CPU performance if it becomes too hot.

Taking 3 different CPUs, similar SSD, Ubuntu, let’s see how OAI performs with OAI PHY unitary tests

General purpose usage: let’s compile the OAI PHY simulators

 N3160
Ubuntu 17.04
i5-4300M
Ubuntu 16.04
i5-6600K
Ubuntu 17.04
Total cpu cost cumulated all cores395 sec146 sec75 sec
Elapsed time216s sec80 sec40 sec

Now, we run OAI phy processing on one core

 N3160i5-4300Mi5-6600K
Detected cpu_freq1.6 GHz
(turbo: 2.2GHz)
2.6 GHz
(Turbo 3.3GHz)
3.5 GHz
(overclock: 4.1GHz)
UE mode: Total PHY proc tx339.031893 us72.814188 us50.013725 us
UE mode: only
ULSCH multiplexing time
20.204291 us7.258590 us5.122038 us
eNB mode: total time483.742061 us 86.693388 us62.064970 us

Conclusion:

N3160, that is recent Atom core architecture (Braswell) is much slower than main stream core architecture.

On main stream CPUs, Intel improved but not that much between Haswell architecture and Skylake.

The difference between the two i5-xxxx is also small because both have AVX2 instructions.

Intel developed more the GPU part between gen 4 (Haswell) and gen 6 (Skylake).

If we take a older CPU like ix-3xxx or before, the difference will be larger for same frequency.

If some people perform similar tests with i7-xxxx or Xeon, they will find out the same values for the same core architecture, no matter the commercial denomination.

Leave a Reply

Your email address will not be published. Required fields are marked *