100 Core Atom
If Intel managed to get the atom's power consumption to half a watt and shrunk the size in half, and put a 100 of the atoms on one chip to make a 100 core atom, Than paired it with 2 ATi Radeon HD 5870's in a gaming console, how would the gaming performance compare to a Xbox 360 and a PS3. The games would be coded to make use of all 100 atom cores and both gpu's.
It would suck pretty badly, you still have lots of serial tasks that you need some raw speed for and the atom will never be able to compete on that level, and no conventional software can ever be coded to make use of more than about 12 threads at the same time, image processing, video editing, and special scientific apps are about the only ones that can make use of a ton of threads as they can be massively parallel applications, but there is a finite limit to how parallel you can make a game, and the slow clock speed and low IPC of the atom would restrict speed significantly. More cores arent super awesome, you still need to be doing a task that has little to no serial restrictions.
What would be the performance be, if you had two identical 3ghz Pentium dual core processors, and the one was clocked in half to 1.5ghz while the other one stayed at 3ghz but had one of it's cores disabled making it a like a uni core. How would the video encoding performance be if the program was optimized for both uni and dual core processors.
The dual core would lose, always. A program cannot be perfectly optimized for a dual core to give you an additive effect of the cores, if it ever encounters a point that is not parallelized because there is a data dependency then the second core will have to wait for a moment until the data comes in then they can both go back to work. You also lose some performance due to the overhead of the thread scheduler and needing to move data between caches. Massively parallel setups are only good for massively parallel tasks, anything that has any serial moments in it, that serial moment will not go any faster regardless of the number of cores.
Well, what could be parallelized?
AI: Imagine something like Starcraft with 1K AI roaming around. Each AI could be on a single core as it's totally independent of the others. Now, many games aren't like an RTS, so there will probably be only 10-20 AI roaming around most of the time.
DX 11: They just made this parallelized somewhat, this could be seperated from other tasks. Should take up a few threads.
Image drawing for GPU's: This could take at least one thread from the processor.
Controls & background tasks: One more thread down
Ethernet usage: Multiplayer games using an onboard port will take a thread.
CPU physics: This could definitely be parallelized between however many processors you want depending on the load. Though, it's no where near as good as a GPU.
Sound: One thread
Hmmm, fill in anything that I've forgotten.
shafe88 said:If one powerful core is better than two low power cores, than why doesn't Intel instead of making a dual core atom just make the atom more powerful while keeping the power usage around the same as the dual atom.
One core isn't always better than two.
Two will allow one core to work on a serial task that takes up the entire core while you could surf the web using the other core.
There's a point where there are too many cores, because not all workloads will be able to use them. You could make loading web pages split into at least 4 cores, but what's the point when our modern x86 procs can easily do this in a second on 1? 4 would just use more power than needed.
shafe88 said:If Intel managed to get the atom's power consumption to half a watt and shrunk the size in half, and put a 100 of the atoms on one chip to make a 100 core atom, Than paired it with 2 ATi Radeon HD 5870's in a gaming console, how would the gaming performance compare to a Xbox 360 and a PS3. The games would be coded to make use of all 100 atom cores and both gpu's.
Intel already did something cloze to that. Not in terms of power usage but something close. Terascale. It used 62.5w of power with 80 cores.
Of course it was put on Intels 65nm process, not current 32nm process. So that may make a major change.
hunter315 said:Massively parallel setups are only good for massively parallel tasks, anything that has any serial moments in it, that serial moment will not go any faster regardless of the number of cores.
It takes 9 months for a woman to produce a baby. If you put 9 women to work on the task, it still takes ... 9 months.
Maybe you could try gaming on Tianhe.Quote:Tianhe-I, Tianhe-1, or TH-1 (天河一号) (Mandarin pronunciation:pinyin: Tiānhé yīhào), in English, "Milky Way (literally, Sky River) Number One", is a supercomputer capable of an Rmax (maximum range) of 2.566 petaFLOPS; that is, over 2½ quadrillion (thousand million million) floating point operations per second, or FLOPS. Located at the National Supercomputing Center in Tianjin, China, it is the fastest computer in the world and one of the few Petascale supercomputers in the world.
It could give you a feel of the gaming you would experience on a 100core monster.
jsc said:It takes 9 months for a woman to produce a baby. If you put 9 women to work on the task, it still takes ... 9 months.
How can that even compare, It's like trying to use two VCR's to record the same two hour TV show, their both recording the same show but it still takes 2 hours to record it.