Personal computing discussed
Moderators: renee, Flying Fox, morphine
Ryhadar wrote:I don't think it's so nonsensical. It's a valid point. Pure, across the board performance improvements aren't really as great as they used to be. On top of that, it's taking longer and longer to get to that next process node so the "just add more transistors because we shrink them" ideology won't work nearly as well in the near future.
So I think the issue is two fold: 1) We're concentrating more on power improvements than pure performance improvements 2) The tricks engineers have been using for years to get more performance out of a chip (higher click speed, smaller node, etc.) are going to hit walls soon. We're going to need some fresh ideas to keep performance going up.
The fact that there ARE ways to continue improving -- new materials such as graphene, or III-IV materials, and new ways of processing (quantum computing, etc) -- means that we haven't hit the pinnacle.
If you wanted to ask "have we reached the pinnacle of Intel Core design?" well, the answer is still no, but it's a heck of a lot closer.
chuckula wrote:Despite what you see in most online benchmarks, the very large majority of software out there that people use every day is *not* heavily parallelized.
Waco wrote:chuckula wrote:Despite what you see in most online benchmarks, the very large majority of software out there that people use every day is *not* heavily parallelized.
Nor does it usually take advantage of specific architectures in any way.
I'd be happy to see the day where I could compile my games/apps on Windows to suit my specific machine instead of using precompiled binaries...but I highly doubt that day will come soon.
ronch wrote:It may also be said that x86 is perhaps one of the most evolved (if not the most evolved) ISAs ever thanks to Intel's Ivy Bridge and upcoming Haswell microarchs (can you also say Bulldozer? hmm..). For comparison, consider Sun/Oracle's SPARC T4, the first chip from Sun/Oracle to implement Out-of-Order integer execution and was introduced only 2 years ago. So when did x86 had OoO? Pentium Pro and Cyrix 6x86 had this feature way back in 1995.
Waco wrote:I'd be happy to see the day where I could compile my games/apps on Windows to suit my specific machine instead of using precompiled binaries...but I highly doubt that day will come soon.
auxy wrote:Ryhadar wrote:I don't think it's so nonsensical. It's a valid point. Pure, across the board performance improvements aren't really as great as they used to be. On top of that, it's taking longer and longer to get to that next process node so the "just add more transistors because we shrink them" ideology won't work nearly as well in the near future.
So I think the issue is two fold: 1) We're concentrating more on power improvements than pure performance improvements 2) The tricks engineers have been using for years to get more performance out of a chip (higher click speed, smaller node, etc.) are going to hit walls soon. We're going to need some fresh ideas to keep performance going up.
It's completely nonsensical. The fact that there ARE ways to continue improving -- new materials such as graphene, or III-IV materials, and new ways of processing (quantum computing, etc) -- means that we haven't hit the pinnacle.
If you wanted to ask "have we reached the pinnacle of Intel Core design?" well, the answer is still no, but it's a heck of a lot closer.
Scrotos wrote:I actually wonder if fabrication technology will eventually allow for a new kind of design? For instance, get some of IBM's integrated water/coolant channels and be able to layer a chip into a 3D monster. I've seen some talk about it a bit but nothing concrete. Would allowing for more interconnects let you come up with a radical design? I don't know enough about chip layout and design to make an intelligent comment about it, but maybe there's some university boffins designing stuff using something beyond AND/OR/NOR/etc gates and they just need some funky new transistor type or fabrication to make it happen.
Speaking of which, where's our dang memristors and flying cars?!?
superjawes wrote:Scrotos wrote:I actually wonder if fabrication technology will eventually allow for a new kind of design? For instance, get some of IBM's integrated water/coolant channels and be able to layer a chip into a 3D monster. I've seen some talk about it a bit but nothing concrete. Would allowing for more interconnects let you come up with a radical design? I don't know enough about chip layout and design to make an intelligent comment about it, but maybe there's some university boffins designing stuff using something beyond AND/OR/NOR/etc gates and they just need some funky new transistor type or fabrication to make it happen.
Speaking of which, where's our dang memristors and flying cars?!?
Chips are very 2D because you're basically acid washing a wafer. You put a mask on so that you end up carving the paths needed, but that process doesn't really lend itself to 3D arrangement.
The other issue is cooling. The transistors used actually don't generate heat in steady state, but they do generate heat when switching, which happens a few billion times a second (a billion per 1 GHz). So even if you could come up with a 3D design, cooling everything equally could be a problem. You mention IBM, but all my relevant Google hits are from five years ago. Doesn't look like anything ever came from it.
And speaking directly to "university boffins," that stuff is already deeper than logic gates in the form of CMOS, and I don't think you're going to get much better than that because, as I mentioned, they only generate heat when switching because they don't draw current (at least not very much) at steady state.