Is there a relation between CPU benchmarks and the time the CPU takes?

[ebelendez]

Hi, does anybody knows if there is a relation between CPU benchmarks and the time the CPU takes?
If my CPU ranks 3,000 and different tasks take about 3 minutes; does that mean that a CPU of rank 9,000 will take about 1 minute for each of these tasks?

If not, is there a way to make an approximate calculation of this type?

Thanks

[David (PassMark)]

There is a rough linear relationship, but it depends a lot on the task.

Two examples,
If the task uses the disk a lot, then a faster CPU won't help much.
If the code used by the task was single threaded, then adding a CPU with more cores won't help.

[ebelendez]

Ok, if the tasks don't use disk and are multi threaded.. is there a way to find this time relation?
Even if is not very accurate, I just want a have an idea of what to expect..

[David (PassMark)]

Even some multi-threaded apps don't scale well.

• For example some databases have semaphores that serialize access to common structures (e.g. during the update of a database index). Read operations might scale well, but not inserts.
• Some other applications depend a lot of the RAM bandwidth so adding cores only works up to a certain point. Performance increase as you add cores will be linear up until a certain point, then be flat.
• Some applications (most in fact) aren't aware of processor groups and NUMA. So their scaling is limited.
• Most applications use a certain amount of RAM per thread. So adding threads can exhaust the available RAM resulting in low CPU utilization as disk swapping occurs. This is especially bad for 32bit apps that never got converted to 64bit.
• Disk & RAM aren't the only bottlenecks. Internet speed, network latency & GPU speed can also be problems, depending on the type of application you have (which you haven't mentioned)
• Some CPUs have hyperthreading (virtual cores). These don't scale nearly as well as full cores. Again it is the workload dependent.
• CPU Turbo clock speeds decrease as you load up more cores. So in a very fundamental way 2 cores aren't twice as fast as 1 core.

But if you want to make vague generalizations, which will usually be wrong, then yes. If you double the CPUMark, you half the processing time.