Report: Roadblock Ahead for Multi-Core Processors
Report: Roadblock Ahead for Multi-Core Processors
Posted 12/08/2008 at 11:30am
| by Paul Lilly
9
Comments
Print
In case you haven't noticed, multi-core processing has taken hold and the race is on to cram more cores onto a single die. But assuming developers can keep up, at some point, chip manufacturers are going to have address a potential major problem that could make adding more cores a useless endeavor. More specifically, a "memory wall" looms large in the not too distant future that, as Jon Stokes from ArsTechnica puts it, could make more than 16 cores pointless.
The problem stems from memory bandwidth not being able to keep pace with faster processors, whether those speed bumps come from a faster frequency or more cores. Put simply, memory is creating a bottleneck and can't feed the processor fast enough, a problem that has existed for some time. Intel and AMD have been able to mask the problem by adding more cache, but doing so doesn't overcome the memory wall, which looks poised to really rear its ugly head as more cores are piled on to new chip packages.
"Engineers at Sandia National Laboratories, in New Mexico, have simulated future high-performance computers containing the 8-core, 16‑core, and 32-core microprocessors that chip makers say are the future of the industry," writes Samuel K. Moore at IEEE Spectrum Online. "The results are distressing. Because of limited memory bandwidth and memory-management schemes that are poorly suited to supercomputers, the performance of these machines would level off or even decline with more cores."
According to the simulation, performance gains level out at 8-cores, with 16-core chips performing no better than a dual-core processor. As the core count increases, the performance starts to decline sharply.
One solution Sandia has proposed is to stack memory chips on top of the processor, something both Intel and IBM have been working on. Some might also remember reading about multi-core memory, a new memory architecture developed by cryptographer Joseph Ashwood. But no matter what the eventual solution, someone's going to have to come with something relatively fast. Sixteen core chips might not be as far off as you think, and the memory wall isn't going to crumble on its own.
Image Credit: Tpan.com
More like this
visibly_stealthy
December 09, 2008 at 6:46am
Intel is making a step in the right direction with integrated memory controllers? I mean....I'm no AMD fanboy but i have a four or five year old AMD Athlon processor that has an integrated memory controller. I think this is one of the main reasons that AMD had slower(clocked) processors than Intel, but performed much faster. However this is not true nowadays with intels high cache sizes that AMD finally started to implement with their latest processors.
In the hands of a master, any object can become a field improvised, lethal weapon.
AndyYankee17
December 09, 2008 at 12:40pm
I didn't know AMD had IMCs
but as far as more work per cycle, that depends a number of factors.
Keith E. Whisman
December 08, 2008 at 11:44am
The problem is with the way memory is installed and used on PC's..
Look at the gobs of memory bandwidth that video cards get. Now compare that to the fastest memory available on a X58board with the highest performance setup available is just a small percentage of what a video card has.
What needs to be done is a different interface. We need to quit looking at ram as something that should come as a module card.
Perhaps we need to go back to the old fashioned way of individual memory square memory chips that you can plug into your motherboard. Or memory can come in the form of a card that looks like a CPU with lots of pins on one side that plugs into a ziff socket. Those pins would offer that bandwidth needed. The chip would be located as close as possible to the CPU.
I get the feeling that the memory makers would resist such changes like the plague. Even if we had DDR5 modules the bandwidth would still be a hinderence because of the form factor.
We need to change the way we look at memory. The way we buy memory. It needs to change in order for computer technology evolve.
I believe this and have believed this for a long time. Just look at your video card and imagine what if your CPU had access to that much memory bandwidth.
Someone is holding our computers back on purpose. Video cards don't have limitations. Why should our motherboards be held back...
Everyone needs to get pissed off. You should be pissed. Get mad. Get upset. The only way we are going to see any change is if we show how pissed we are. We are the consumers and consumers drive technology developement. So get mad.
AndyYankee17
December 09, 2008 at 12:43pm
ziff would probably be nice, that's probably the reason why newer RAM types have more pins.
as for closer physical location that would probobably make virtually no difference, placement based on better airflow would be better
QUINTIX256
December 08, 2008 at 3:02pm
"is just a small percentage of what a video card has."
A typical setup with a pair of 333mhz ddr2 sticks has a theoretical maximum throughput of over 10GB/s.
In the xbox 360 the maximum theoretical throughput between the GPU and the (shared) memory is around 22 GB/s.In a radeon 4850 it is around 63 GB/s; or roughly 5 times the bandwidth available from dual-channel 400mhz ddr2. It is very significant, but I wouldn't go so far as calling 19% a small percentage.
Also, the GPU and CPU communicate over a relatively slow PCI express pipe (8GB/s). The importance is not so much the raw bandwidth between proc and ram as the fact that GPU ram usage and CPU ram usage are isolated and don't interfere with one another.
Considering that JEDEC is responsible for keeping both DDR and GDDR standards, and that memory chips for CPU and GPU are often designed/manufactured by the same companies (say Qimonda), I would doubt any foot dragging on the part of memory manufacturers.
nekollx
December 08, 2008 at 12:18pm
small niggle bit We will probably never see DDR more then 3, its just not pratical. What pratical difference is there between DDR4 and DDR2 x2
More then likly (if things don't change as you suggested) it would be motherboards with DDR2+3
AndyYankee17
December 08, 2008 at 12:26pm
DDR4 vs DD2 x2?
no offense but you need to study how memory works.
and memory bandwidth is lagging behind although at least intel is heading towards the right direction with integrated memory controllers
nekollx
December 08, 2008 at 12:31pm
from a logical standpoint 4 synroyness rows of RAM isnt really all that better from Two Asynronus rows. But more then that you have to explain to the Consumer why they need 4, 5, 6, 7, 8 Pieces of ram just for it to be "efficient"
Shalbatana
December 08, 2008 at 11:40am
By the time they update/create new, re-compile all the programs to use 16 cores, assuming I need 16 cores, assuming I can afford 16 cores, assuming they have re-designed mobos to hold 16 cores, and if GPU's haven't replaced 16 cores... I think they will have sushed out the problem by then.
_______________________________
"There's no time like the future."
Connect with MaximumPC
Featured Content
We introduce Intel's latest class of tablet killers and review 4 of the first Ultra...
Where have all the memorable videogame enemies gone?
This month's issue
Feature
11 Hardware Hacks
Feature
Overclock Your CPU
How-To
Supercharge Your Smartphone
Build It
A Powerful $830 Rig
Most Commented Articles
Latest Max PC Tweets
- maximumpc: Multi-Monitor Usage On The Rise http://t.co/eWfJfCEG10 hours 9 min ago
- maximumpc: Windows 8 Consumer Preview Coming On February 29 http://t.co/3Nd04FdR15 hours 57 min ago
- maximumpc: Can RIM *ever* catch a break? Not today! App World numbers not as great as previously reported http://t.co/3ERRhr311 day 11 hours ago
