Reversing the history of computer architecture
|
Prev: MAKE ALL IT CERTIFICATIONS - A PIECE OF CAKE
Next: GL_MAX_PROGRAM_EXEC_INSTRUCTIONS_NV limitation "per pixel" or "per frame" ?
From: "Andy "Krazy" Glew" on 23 Oct 2009 00:47 Andy "Krazy" Glew wrote: > Some of them think that the advent of Larrabee and Atom show that OOO is > a dead end. They think that we are resetting to simple P5-era in-order > machines. More, reversing evolution: retreating from Pentium 4 > "fireball", backing out of OOO. Some think that we will never go back to > OOO. Me, I think that we are resetting. I think of it as a sawtooth > wave: backing out a bit, but probably advancing to dynamic techniques in > a few years. By the way, when I reread this I am reminded that I myself have posted about this possibility. If I make the following assumptions: * Transistors are free * But power is the most important thing * and you have a lot of parallelism as is true of some supercomputer workloads (and even virtual reality graphics for the home) Then I think that you can reasonably extrapolate that the best computer architecture is MIMD, with the simplest possible, non-pipelined, blocking on a cache miss, processor cores. Reason: the slightest bit of microarchitecture smarts - pipelining, multithreading - costs. And wastes power. These technologies exist to make more efficient use of a limied number of transistors, but if transistors are free... This extrapolation being valid in certain circumstances, there are situations where it is not: First, implicit in the "transistors are free" is that leakage can be neglected - e.g. by making the transistors slow, slow, slow, and by using parallelism to make up for the slowness. But if leakage cannot be eliminated from consideration, then you add back just enough microarchitecture smarts, just enough pipelining, that the performance gained back reduces the processor count by enough to balance the power lost due to the transistors for the smarts. Second, communications has costs. If you have to double the wires to get the extra, slow, unpipelined processors, it may be worthwhile adding some uarch smarts back. Actually, it's not "double the wires", but "double the bits*wirelength total", but you get the idea. And this is workload dependent, and communications technology dependent. Go optical, and you have costs, but the long distance costs may be less, encouraging more, simpler, processors. Finally, and for the moment the most important: transistors are not free. Yet. We do uarch smarts, stuff like multithreading, because it is more cost effective than simply adding cores.
From: Mayan Moudgill on 23 Oct 2009 07:59 Andy "Krazy" Glew wrote: > First, implicit in the "transistors are free" is that leakage can be > neglected - e.g. by making the transistors slow, slow, slow, and by > using parallelism to make up for the slowness Data point: I suspect typical foundry low-leakage proceses are about one generation behind the G process in transistor speed (i.e. 90G ~ 65LP), with insignificant leakage. I suspect that IBM/Intel gain another 1/2-1 generation's worth of speed for their parts by tuning the process, playing with the corners and sorting (though you can probably do the same with a foundry, if you have high enough volumes).
From: Mayan Moudgill on 23 Oct 2009 08:05 Andy "Krazy" Glew wrote: > Go > optical, and you have costs, but the long distance costs may be less, > encouraging more, simpler, processors. > Optical? In a commodity process? For high speed communication? Unless someone has actually managed to get a standard-process-compatible laser, not going to happen. Costs are too high. I was excited by the iron implant stuff and strained silicon but I don't think they ever made it out of the laboratory. Anyone know enough to comment? An alternate transport mechanism (when using flip-chip) is to drive the signal out to the carrier (specially if it's ceramic) and then use that as the high-speed long-haul layer. [BTW: I'm not sure if its possible to use diodes to do high speed optical communication; anyone know?]
From: Del Cecchi on 23 Oct 2009 16:11 "Mayan Moudgill" <mayan(a)bestweb.net> wrote in message news:IpqdnUOUWcd0BnzXnZ2dnUVZ_oydnZ2d(a)bestweb.net... > Andy "Krazy" Glew wrote: > >> Go optical, and you have costs, but the long distance costs may be >> less, encouraging more, simpler, processors. >> > > Optical? In a commodity process? For high speed communication? > > Unless someone has actually managed to get a > standard-process-compatible laser, not going to happen. Costs are > too high. I was excited by the iron implant stuff and strained > silicon but I don't think they ever made it out of the laboratory. > Anyone know enough to comment? > > An alternate transport mechanism (when using flip-chip) is to drive > the signal out to the carrier (specially if it's ceramic) and then > use that as the high-speed long-haul layer. > > [BTW: I'm not sure if its possible to use diodes to do high speed > optical communication; anyone know?] Not silicon diodes for transmission. As for using the ceramic, you are about 30 years late. see TCM for example. IBM even went to some expense to upgrade the material from Alumina to Glass Ceramic to improve propagation. Also LEDs are pretty lacking in bandwidth compared to lasers. But since silicon is indirect band gap the point is moot. Pipelining wasn't driven by lack of transistors but by frequency. It didn't come in until transistors got sort of cheap. (cheap is in the eye of the beholder). del
From: nmm1 on 23 Oct 2009 16:38
In article <IpqdnUOUWcd0BnzXnZ2dnUVZ_oydnZ2d(a)bestweb.net>, Mayan Moudgill <mayan(a)bestweb.net> wrote: >Andy "Krazy" Glew wrote: > >> Go >> optical, and you have costs, but the long distance costs may be less, >> encouraging more, simpler, processors. > >Optical? In a commodity process? For high speed communication? > > Unless someone has actually managed to get a >standard-process-compatible laser, not going to happen. Costs are too >high. I was excited by the iron implant stuff and strained silicon but I >don't think they ever made it out of the laboratory. Anyone know enough >to comment? Not on that. But the fact that serious money is still going into research in this area, despite decades of next to no progress, indicates that the manufacturers are very aware of the potential. The last plausible idea I heard was lasers pumping in to silicon optical switches, but that one seems to have gone awfully quiet. Regards, Nick Maclaren. |