HP ProBook 430 G5 Mini-Review

We bought a HP ProBook 430 G5, Core i5-8250U, 8GB RAM, 256GB SSD (3KY85EA#ABD) for EUR 761. The HP Probook 430 G5 has a lot of the features that I look for: What could be better: The BIOS (use F10 to enter, F9 for the boot selection) contains no "enthusiast"-oriented options, but instead business-oriented options (e.g., scheduled power-on, or saving the BIOS settings to and restoring them from a USB stick), and also allows to disable the camera and the microphone.

We ran Ubuntu 17.04 and 16.04.3 from a USB stick, and tested various functions:

One interesting aspect was the clock frequencies of the CPU under various loads, with mains power and when battery-powered (the BIOS says that Turbo is disabled when on battery (that can be turned off), but the results below contradict this). We generated load with different numbers of yes >/dev/null processes (integer-only load), and found the following clock frequencies (with some occasional funny results):

     MHz   MHz
load mains battery
  1  3400  3400
  2  3000  3000
  3  2700  2540
  4  2400  2300
  5  2300  2300
  6  2200  2300
  7  2200  2200
  8  2100  2100

Note that the Core-8250U has 4 cores and 8 threads, so load 5-8 uses just hyperthreading on the same 4 cores.

The clock reaction to increasing or reducing the load was usually within 1s, so it is probably power-target-based or load-based, not temperature-based. We did these experiments for at least 10 minutes, but the clock rates were apparently not affected by the temperature buildup for this workload. The fan, which was off (or at least inaudible) at idle did spin up after the load (even just load 1) warmed the CPU up. Spinning up and down was gradual and not very loud, so not as annoying as the fan on my Lenovo E130.

Another performance test was to perform 1000x1000 double-precision matrix multiplication using libopenblas with varying numbers of threads, performing 100 runs between stretches of idleness.

threads MHz   GFlops
  1     3084  25.8
  2     3360  39.6
  3     3227  43.8
  4     3099  44.6
  5     3066  40.0     
  6     3009  40.6
  7     2957  41.4
  8     2892  41.8
It seems that the performance is quite high for several seconds even for multi-threaded AVX-using applications such as this one. After a longer load of this kind, the performance with four threads drops to 16-22GFlops.

The relatively low gains when going from 2 to more threads despite relatively high clock rates makes me suspect that the benchmark becomes memory-bound in that setting (each of the three 1000x1000 matrices involved consumes 8MB, and this CPU has 6MB L3 cache), and all these additional threads do is increase the proportion of time that they are waiting for data from memory.

Anton Ertl