I Jedi

January 16, 2009 at 10:32am

There was a similar article about a month or two ago on MaximumPC in which they described this very exact problem, too. They'll come up with something eventually. Although, I can't imagine why someone would want or need 16 core in one processor altogether... besides the ability to run huge math calculations or server based stuff.

StahnAileron

January 15, 2009 at 11:25pm

Most supercomputers have hundreds, if not thousands of cores, yes...But They're also composed of dozens or hundreds of NODES, each which can contain several individual computers which have their own dedicated systems. A HPC maybe have the aggregate equivalent of several terabytes/sec of Memory bandwidth, but it's NOT a collective bandwidth ALL cores can access. Each individual Core/CPU has only a small fraction of the aggregate bandwdith available to it.

Saying a HPC has a lot of memory bandwidth is like saying it has a very high speed reading becuase you added all the individual core speeds together. (2 x 4GHz cores does NOT equal the equivalent of an 8GHz single core.)

Cache is based on SRAM (far as I recall. Correct me if I'm wrong anywhere. Always up for leanring more), and that's both expensive as hell and ineffifcient in terms of silicon use. (DRAM is preferred for high density RAM use because of how much denser you can pack it together compared to SRAM...Assuming similar fab size, IE 65nm.)

I don't see cache-type RAM being used as main system RAM any time soon...Not unless cost and density have been addressed. (Current mainstream SRAM is 6T, no? I don't recall if lower T-count SRAM designs have been implemented in practice yet...)

And with SRAM, there's the problem of physical distance, especially at high frequencies. Signals can only travel so far in the billionths of a second a modern CPU's clock rate can hit. SRAM can run pretty quick, but once you move it a good way from the CPU, you negate it's speed advantage because the signal starts to lag. Then your Price vs Performance ratio would suffer, I'd think.

One of the problems I see ii the overhead needed to ochestrate the distro of data on massively multicore systems. I mean, if the OS can't take advantage of all the cores available to it and/or efficiently dole out work to the cores, what's the point?

Of course there's the software application side as well...(It'll be the day when something like Notepad can make use of multiple cores, LoL)

Maybe they can start designing consumer systems with dedicated Memory controllers for EACH core or something...Maybe like how the CPU cache designs are used, but at the system RAM level using DRAM. (L1 RAM with dedicated connection to a core, Shared L2 RAM for full system use.)

It'll be interesting to see how they solve the memory bandwidth problem though. They got past the perfomance versus Core speed/thermal issue by going to Multi-core...Let's see where they go with RAM.

mikemckay

January 15, 2009 at 10:05pm

i just have to throw this out there.....havent we seen this all before?

i meen now you can get single video cards with many hundreds of cores (or stream processors...whatever) and more than 100GB/s bandwidth to the memory and THEY still seem to scale well......are we supposed to assume that system memory will never advance past triple channel ddr3?

those tests just seem like they might be a little unrealistic

devin3627

January 15, 2009 at 3:06pm

Is that talking about motherboard/ddr3 fsb? What if you had 3000mhz ram.....? There's gotta be a instruction code around this that uses L3 cache, externally?

AndyYankee17

January 15, 2009 at 12:46pm

depends heavily on what you're trying to do, realtime things like games are hitting a block with multicores but nonrealtime things like encoding/decoding, rendering, compiling really fly with multiple cores

decapitor

January 15, 2009 at 12:25pm

That's a pretty vague argument, especially since clusters and supercomputers have been rocking thousands of cores for quite some time now and helping the scientific community crunch data at incredible speeds.  I just xgridded 5 8 core mac pros and my mpi code is significantly faster than on one machine.  I mean of course no matter how well optimized your code you are going to have diminishing returns as you add cores, but the performance always increases.  Just put me in the "doubters" bin for now.

QUINTIX256

January 15, 2009 at 12:21pm

'report: roadblock ahead for multicore processors'

Also, "key algorithms for deriving knowledge from large data sets" seems to me like a lot of cache hit misses. If the cpu is doing a whole lot more time waiting for the cache to be swaped out than actual data crunching, than yes, it is much more of a measure of memory bandwidth than a measure of processor performance.

You can have your recession. I'm not participating. (vs The economy is suffering, let's starve it!)