One peculiar consequence of fat CPUs with enormous core counts is that measuring CPU load in percentage of allocated time slots - as in, most monitoring software does now - is kind of useless.
For example, when talking about a CPU with 64 logical cores, one process utilising two cores to the full will only show up as 3% load, and yet this might just be enough to ramp up the cooling system and make some noise. Meanwhile, a bunch of small loads spread across the cores will add up to the same 3% and yet the overall load will be lower, your fans will stay quiet. If you throw dynamic clocking into the mix, the entire idea of counting allocated time slots goes to shit. 50% load on a core at 2GHz and 50% load on a core at 5GHz are entirely different things. Therefore, this metric is most useless any way you look at it.
The better idea, as I see it today, would be to RETVRN to tradition and measure load in... Watts against the maximum TDP that your CPU can draw. A mostly idling CPU at low clock wouldn't draw much power, whereas a CPU at high load would draw as much as possible. Number of processes or cores loaded doesn't matter here, and the number is pretty easy to grasp.
@splitshockvirus i didn't even mention how different instruction sets affect performance and/or power draw. On Intel CPUs, a process calculating md5 for an infinite data set at boosted clock rate would count as 100% load, but it won't draw the maximum possible TDP. But if you utilise AVX512, your power budget needs to double or even triple compared to regular code. Meanwhile, the CPU clock would plummet at the same time because AVX512 works only at reduced frequencies.
Add comment