The problem is not Moore's Law. Yes, we keep doubling the number of transistors in the same space (or same number of transistors for half the space). The big problem is power. Its refered to as the power wall. We just can't get enough power on to the chip to power all the transistors. Another consideration too is the memory wall. Once more also the speed at which CPU's frequencies have increased has not matched that of Memory interface speeds.
But frequency is not a very good judge of performance. As you say, more cores should mean more speed in terms of performance. Unfortunately that is not true...
I know what you're saying, and you're right, mostly.
For stuff that can't be parallelized, for the past few years, speed has been going up much slower for the reasons you mention. For stuff that can (quite a bit, graphics being the obvious example), more cores = more speed, and for that, performance has been increasing faster than moore's law.
The problems you mention (power, memory speeds), that's only a barrier if one stays with current designs of semiconductors and related technologies... but, that's an evolving field, and will change, likely bringing with it dramatic performance increases. Looking back 20 years from now at a chart of CPU performance over time, we'll probably see this line ramping up, then almost a plateau (which we're in right now) then an abrupt vertical jump continuing or superceding the previous line. IMO anyway :)
I still think you are being overly optimistic. Power is a much bigger issue than you make it out to be. Did you know that current cpu's have a higher power density compared to nuclear reactors?
Sure.. but CPUs are not nuclear reactors (another field which will soon significantly change).
Stuff sometimes seems not to change.. but it doesn't mean it's not changing (development etc then ramping up to production) behind the scenes for the most part.. but changing it is.
Time will tell who is right.. but Intel have their entire business at stake in moving forward. There's obvious areas to explore, and they're exploring it (and I'm sure what's public is only a part at what they're looking at). They'll succeed.
(well, they sort of did, but the only reason you don't see larabee processors is because they couldn't make the GPU side cost effective enough for the given performance.. be very sure though that the tech still exists and is being developed, and will appear at some point)
![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
![[identity profile]](https://www.dreamwidth.org/img/silk/identity/openid.png){kind=small}
no subject
Date: Tuesday, 23 March 2010 13:02 (UTC)But frequency is not a very good judge of performance. As you say, more cores should mean more speed in terms of performance. Unfortunately that is not true...
no subject
Date: Tuesday, 23 March 2010 15:19 (UTC)For stuff that can't be parallelized, for the past few years, speed has been going up much slower for the reasons you mention. For stuff that can (quite a bit, graphics being the obvious example), more cores = more speed, and for that, performance has been increasing faster than moore's law.
The problems you mention (power, memory speeds), that's only a barrier if one stays with current designs of semiconductors and related technologies... but, that's an evolving field, and will change, likely bringing with it dramatic performance increases. Looking back 20 years from now at a chart of CPU performance over time, we'll probably see this line ramping up, then almost a plateau (which we're in right now) then an abrupt vertical jump continuing or superceding the previous line. IMO anyway :)
no subject
Date: Tuesday, 23 March 2010 15:21 (UTC)no subject
Date: Tuesday, 23 March 2010 15:25 (UTC)Stuff sometimes seems not to change.. but it doesn't mean it's not changing (development etc then ramping up to production) behind the scenes for the most part.. but changing it is.
Time will tell who is right.. but Intel have their entire business at stake in moving forward. There's obvious areas to explore, and they're exploring it (and I'm sure what's public is only a part at what they're looking at). They'll succeed.
no subject
Date: Wednesday, 24 March 2010 07:26 (UTC)no subject
Date: Wednesday, 24 March 2010 13:11 (UTC)(well, they sort of did, but the only reason you don't see larabee processors is because they couldn't make the GPU side cost effective enough for the given performance.. be very sure though that the tech still exists and is being developed, and will appear at some point)