forget 64bit.... here comes the 64CORE cpu!

[BrianKitts]

http://arstechnica.com/articles/paedia/cpu/MIT-startup-raises-multicore-bar-with-new-64-core-CPU.ars

[AJLynn]

Pretty cool indeed.

 

Now for the software to catch up... this one-cycle-to-propagate-one-step system makes it entirely not like current multithreaded computing.

[JamesTaylor]

can someone translate that link into paraphrased layman terms for us less intelligent people please!!

 

all sounds good tho!!!

[BrianKitts]

can someone translate that link into paraphrased layman terms for us less intelligent people please!!

 

in the words of eric cartman.... "I don't know, but it sounds pretty sweeeeeeeet!"

[AJLynn]

Okay. These guys have made a 64-core CPU. It's a lot like the one Intel has under development, but it's not available for mainstream PC use and for the forseeable future it will be a server chip.

 

When you have that many CPUs, internal communication and coordination become very difficult. System buses can't cope with 64 CPUs. So instead of making each CPU separately available to the PC over the system bus (as in a normal multi-core system) you put them in a grid (hence the "tile" name) and put some of the ones on the edges of the grid on the bus...

 

This is analagous to molecular bonding, remember the models with the balls and sticks where the balls are atoms and the sticks are bonds and the number of holes in the balls represent how many bonds the atom can make. One core of this type is a ball with 4 holes, and if you put a bunch in a grid you can attach each one to four others. The ones on the edges have holes that aren't being used, so this is where you can attach other stuff like networking and PCI-Express controllers and access to main system memory. The cores all run on the same clock, say 800MHz, and in each cycle a core can do calculation and can pass some amount of data over the links, which is how the coordination works - this is different from the current chips where each core is on the system bus and functions independently.

 

So now some interesting new problems start to come up. Your code needs to be run in parallel in different ways from how the current multithreaded code runs. It needs to account for the closer cores being able to cooperate more efficiently and the lag in sending data to cores far from the edges. This requires, minimally, recompiling of the operating system and the software, and to get the most out of it, some changes in the way the code is written.

[Arnold Sher]

well, i am glad it's all cleared up... Lunch anyone?????

[JamesTaylor]

Andrew, so should these chips be stupidly fast??? once the software catches up that is??

 

i presume at present non of our standard apps such as CAD, Viz/Max, Vray, photoshop, after effects, combustion even windows etc. etc. will be able to utilise the power of these chips??

 

Finally is it likely that applications will be redesigned/written (perhaps over a few releases) to utilise the methods of transferring information across this chip or will we see more traditional multi core cpus continue to develop and be the processors of choice?? Afterall in the space 2 years we've moved from single core machines to 8core machines here - another 2 years at that rate would see us at 64 cores also????

[pradipta]

awesome, should have insane power!!!

[AJLynn]

Andrew, so should these chips be stupidly fast??? once the software catches up that is??

 

i presume at present non of our standard apps such as CAD, Viz/Max, Vray, photoshop, after effects, combustion even windows etc. etc. will be able to utilise the power of these chips??

 

Finally is it likely that applications will be redesigned/written (perhaps over a few releases) to utilise the methods of transferring information across this chip or will we see more traditional multi core cpus continue to develop and be the processors of choice?? Afterall in the space 2 years we've moved from single core machines to 8core machines here - another 2 years at that rate would see us at 64 cores also????

 

Stupid fast, yes. The example they gave was simultaneous real time encoding of 10 broadcast-quality HD streams. That's one of their target markets, the other is servers, maybe rendering comes reasonably soon...