arstechnica.com — The "memory wall" was been looming in the path of computer performance advancements well before clockspeed increases hit the "power wall." In fact, in the many-core era, we may hit the memory wall quicker than we expect.
Dec 7, 2008 View in Crawl 4
squigglethecowDec 8, 2008
why did you get dugg down for that?And why am I always asking that question?!
paranoydandroidDec 8, 2008
I hate to get all mathematician here, but this is really bothering me.Your use of "hypercube" instead of "cube" or "cquare" indicates you're talking about some n-dimensional analogue of a cube (where n > 3), but it would be impossible to have that shape on a motherboard (or anywhere in existence for that matter).So really they have a _projection_ onto 3 dimensions of some n-dimensional hypercube (presumably a 4-dimensional hypercube).
sanosuke001Dec 8, 2008
His original quote also had a piece about "on current hardware" attached to it. I don't remember the exact quote and don't care to look it up; but it wasn't "nobody will ever need more than 128K RAM" and that was it. He did recognize that in the future, it would be useful. But on current hardware, it wouldn't be.Also, it isn't how memory is addressed that is the issue. We can address more memory than we have available on a 64-bit system. It's how cores interact with memory that is the problem. Legacy x86 instructions are the biggest problem.
sanosuke001Dec 8, 2008
The problem is that x86 has been around too long and everyone has too much money invested into it. It would be prohibitively expensive to switch to something new. Though, they could always create an x86 emulator for whatever they create so legacy software would still work. But what's the point? AMD and Intel are making wad-loads of cash on x86. However, they would have to decide together to switch to a new architecture for it to even be remotely possible.
xsledgewickDec 8, 2008
Yeah, but... 2TB? I mean come on...
topher06Dec 8, 2008
When I can rip a DVD in under 2 minutes, then I will claim there is enough cores for home use.
evilgourmetDec 8, 2008
you mean pointless with current software, that is not written for parallel processing -- and current architecture that is not parallel.With new hardware arch and new software written for the system, not overwritten 16 >> 32 >> 64, the system would run fine, and fast.Poor dumb bastards will have to run the crappy (insert 1990's software name here) in an emulator instead of natively -- deal with it.
fipiDec 9, 2008
true. but it's not even a projection. it's just any structure that is isomorphic to a hypercube. in general, in networking and parallel computing terminology, we use geometrical terms only to describe the edge structure, rather than the physical structure.