Author: Hristo
Date: 08:53:05 11/03/05
Go up one level in this thread
On November 03, 2005 at 03:57:07, Dan Andersson wrote: >On November 02, 2005 at 21:42:18, Hristo wrote: > >>On November 02, 2005 at 16:19:03, John Dillard wrote: >> >>>On November 02, 2005 at 15:34:30, Joshua Shriver wrote: >>> >>>>http://www.apple.com/powermac/ >>>> >>>>nice :) would make a good quad system. >>>> >>>>-Josh >>> >>> >>>They're making a quad system. There's not other system on the market today, >>>super computer or otherwise, that can process as many gigaflops of info as the >>>dual core G5. I just wonder if any of the chess programs will benefit from this >>>power? >> >>John, >>I love Apple computers. In fact I'm writing this on my favorite PB 17" (OS-X >>10.4.3). The other fact is that at work I use Opterons (three different >>systemsin my office, all of them dual CPU) and those systems are able to match >>or destroy (in some cases) the newer Macs (which we also have at work). >>The only time a Mac wins (against Intel ot AMD) is when you can fit your problem >>solution into Altivec and then spend some time optimizing it, which we have done >>in a few cases in the domain of signal analysis. Outside of the Altivec-unit the >>Macs are not going to win against AMD. >>Memory access latency is the killer for many apps, not the memory access >>throughput. In this sense most chess programs are limited by random access > > The PPC/G5 problem concerning integer code isn't primarily due to memory >latencies. It is a design issue. It almost always is. :-) > First of all IBM is using a building block approach when designing the CPUs >that forces a two cycle latency on integer instructions. Latency here, latency there, latency everywhere. Branchy-code which is using only float (double) type of instructions and data types also gets hammered; scalar mode. Once things get running in a vectorized mode, then it doesn't really matter whether you are doing integer or float ... it's all good. :-) Anyway, the whole thing (G5) seems to be designed to have huge throughput to memory and then to rely on considerable amounts of cache on the CPU (to reduce the cost of going to main memory), which the G5 doesn't have. I guess the part has to be cheap too. :-) > That results in a cost >in tight branchy loops at mispredicts. The upside is that it siplifies design >and layout. > Secondly the dispatch and issuing of the micro operations have some strict >limitations on packing that reduces the practical IPC. But it lowers the >complexity and power consumption. The power consumption is apparently a tricky bit. I'm not sure about that but G5 probably consumes more power than a corresponding Opteron. Regards, Hristo > >MvH Dan Andersson > >>latencies and not be sheer throughput (as it is needed in video or signal >>processing). The dual-core G5s are not going to win the contest so easily >>against AMD in particular. (In fact I would pay extra to get OS-X run on an AMD >>processor) >> >>Anyway, >>enjoy your Mac for what it is, the best computer experience you can have today, >>and not for what it isn't the best chess playing computer in the world. If >>someone spent the time to translate (port) their chess algorithms to Altivec (if >>this is even possible) then your assumption _might_ have some merit. Until then, >>enjoy your computer and what you can do with it. :-) >> >>Regards, >>Hristo
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.