2 Cores Vs. 4 Cores

[Umbrarian]

Counting no-ops, yes. But hyperthreading turns those no-ops into useful work.

What I am saying is if you are running at 98%, most you can get is 2%, not 30%. For example when I am running Handbrake all cores are running at over 99%. HT adds almost nothing in this example.

 

[robeddie]

The quickest answer is... what are you doing with your computer?
If your tasks involve a lot of CPU cores, then a quad-core will benefit you more. If you don't need to utilize more cores, a dual-core CPU will be fine.

Video editing and encoding is a big one. So its gonna use all 4 cores, real or 'fake'. The 4 'real' cores in the imac, and it'll also use the two 'real' cores in the macbook pro as well as the two 'fake' ones created by hyperthreading. So I think it'll max out the capability of the processors either way. The encoding program will 'see' 4 cores either way, so thats why I asked my original question, which is whether theres any reason to think the 4 'real' cores is any different than 2+2, given the fact they score the same on geekbench.

 

[jerwin]

What I am saying is if you are running at 98%, most you can get is 2%, not 30%. For example when I am running Handbrake all cores are running at over 99%. HT adds almost nothing in this example.

There are ways to toggle hyperthreading on and off so you can benchmark handbrake under both conditions. Of course, if handbrake can only use 4 threads, it won't matter if you have 8 logical cores..

As for cpuload, has it ever occurred to you that those programs don't have access to the necessary data to give a proper estimate? (This site claims that while SPARC and POWER CPUs have "peephole registers" for this purpose while intel CPUs do not.)

From what I understand of the intel core it can execute four instructions per clock cycle, provided that it isn't starved of data, and the instructions are perfectly matched to the resources available in the core-- e.g. two integer and two floating point operations. By letting two threads share the core, it's more likely that this theoretical performance limit can be attained.

 

[Umbrarian]

There are ways to toggle hyperthreading on and off so you can benchmark handbrake under both conditions. Of course, if handbrake can only use 4 threads, it won't matter if you have 8 logical cores..

As for cpuload, has it ever occurred to you that those programs don't have access to the necessary data to give a proper estimate? (This site claims that while SPARC and POWER CPUs have "peephole registers" for this purpose while intel CPUs do not.)

From what I understand of the intel core it can execute four instructions per clock cycle, provided that it isn't starved of data, and the instructions are perfectly matched to the resources available in the core-- e.g. two integer and two floating point operations. By letting two threads share the core, it's more likely that this theoretical performance limit can be attained.

My biggest system is 10 cores and it uses all of them.

Yes it that has occurred to me as I was an engineer at Sun when we pioneered this technology on SPARC. It works well for overall performance of large interactive or TP loads (stuff where there might be 1000s of context switches per second) for but offers little on CPU intensive programs such as HPC.

Most personal computers are single user and do not have enough stuff running simultaneously to take much use of HT. Server based stuff is much more suited.

Of course as you said, anyone can run a given benchmark such as HandBrake with/without HT and see how performance changes and decide for themselves. I generally turn it off myself.

 

[jerwin]

Yes it that has occurred to me as I was an engineer at Sun when we pioneered this technology on SPARC. It works well for overall performance of large interactive or TP loads (stuff where there might be 1000s of context switches per second) for but offers little on CPU intensive programs such as HPC.

Fair enough. I'll have to glance at the hpc literature, then.