8 Physical cores vs 4 Physical 8 Threads?

[beginner99]

I would have choosen the i7. Would overclock better and is faster per core. Probably a close race. i7 is also more power efficient and if you run it 24/7 at 100% that will quickly become a factor.

 

[HAL9000]

I would have choosen the i7. Would overclock better and is faster per core. Probably a close race. i7 is also more power efficient and if you run it 24/7 at 100% that will quickly become a factor.

Not that fussed about power consumption, I do like having 8 cores aswell, plus it cost a lot less than a 980x would have set me back.

 

[senseamp]

Hyperthreading is splitting functional units of a core between two threads. Good if you have two different applications that want to use different functional units.
It's like if you have a men's room and women's room. Good if a two people of different gender want to use restroom at same time, but if there are two guys who want to use it, they'll have to serialize, and if it's two women, well, forget it.
Presumably,video encoding is going to split the video stream into segments for each core to process, but it's going to run similar code in all threads and thus contend for same functional units. In this case you are better having two mens rooms and two womens rooms, even if utilization is going to be lower.

 

[Makaveli]

Good Post

and if it's two women, well, forget it.

This was my favourite part

 

[Denithor]

I have always visualized HT as simply a hardware-level scheduling tool that allows each core to be fully utilized by arranging work in a more efficient manner. Doesn't help in situations where you have enough work to keep all cores busy though, more cores takes the easy win in those cases.

That, of course, assumes an apples-to-apples comparison of cores.

 

[Cogman]

Hyperthreading is splitting functional units of a core between two threads. Good if you have two different applications that want to use different functional units.
It's like if you have a men's room and women's room. Good if a two people of different gender want to use restroom at same time, but if there are two guys who want to use it, they'll have to serialize, and if it's two women, well, forget it.
Presumably,video encoding is going to split the video stream into segments for each core to process, but it's going to run similar code in all threads and thus contend for same functional units. In this case you are better having two mens rooms and two womens rooms, even if utilization is going to be lower.

Not necessarily true (about the video encoders). Video encoders have a lot of integer and floating point math. So the effectiveness of hyperthreading is really going to be a "Hit miss" sort of situation. rather then always being a benefit. In other words, if one thread just so happens to be using the APU and the other the FPU, then hyperthreading will see a nice boost.

Heck, even if they are doing the same sort of instructions (EG, both are doing integer multiplication) there is still a possibility that the parts required for multiplication are open for 2 threads. (in fact, where the instruction is highly pipelined it is likely)

 

[Dufus]

Are OS's generaly suppose to use PCore first then thred onward to the HT/LCore or is just random and depend on the programming/er?

That's a good question. With HT each physical core has 2 threads** which are seen as logical CPUs. Both threads are equal and it shouldn't matter which of the two threads you run on performance wise. A performance hit can however come from running on both threads of the same physical core at the same time.

Here's some results with an i7-860 with HT enabled and W7HP64. Balanced power plan used.

4 physical cores ( Core 0 to Core 3 ) each with 2 threads.

1) This is the usual run of Linx using 8 threads, 33GFlops. Core parking is enabled and defaults mean at least one thread per physical core should be active. This has the effect of the first 4 software threads usually being assigned to different physical cores unless the logical cores are considered overloaded.

2) Linx is set to use 4 threads and because affinity is set to "all", the OS gets to choose which logical CPUs they will run on. They may swap threads when context switching occurs. Well in my case the scheduling didn't go too well with the software threads running mainly on just physical core 0 and core 1 and a little on core 3. Result, 24-28GFlops.

3) Still using 4 threads, affinity has been set to use 1 thread per physical core. This gives a much better result, 37GFlops.

4) Now there seems to be a bug with the core parking algo'. Selecting affinity of the 4 parked cores results in only one thread being un-parked and used resulting in terrible performance. BTW I did not disable turbo for these tests so that 1 core was probably using 26x multi against 4 cores using 22x. I should redo it with fixed multi. Naah, I'm not that keen.

So in my case I guess the answer to your question would be errm... yes, to everything lol !

** Possibly more than 2 threads per physical core may become available in the future.