Petz, you have a point WRT price/performance of high-end workstations, especially considering that you are underestimating the cost of a dual-Itanium system. (Imagine that!) But sometimes you just need all the performance you can get out of a dual-CPU workstation regardless of cost.
Besides, aren't a lot of workstation apps, such as 3D Studio MAX, optimized for two processors? My own work requires using farms of dual-CPU computers running simulations. Increasing the number of CPUs per computer isn't going to make any difference in performance, only the number of jobs that can be run at a time (and even that is limited by the huge memory requirements of each job). Adding more systems is the same. Neither method actually reduces the actual time it takes to run a single job.
<Is there some reason that database transaction code is easily parallelized into 3 instructions per cycle, but SPECint isn't?>
It's not that. Take a look at the leaders in TPC-C for non-clustered quad-CPU systems:
tpc.org
The top two in the 4-way category are based on Pentium III Xeon at 900 MHz. This despite the lack of 64-bit addressing and the fact that all four CPUs have to share a 100 MHz bus. As for the 8-way category, although the P3 Xeon systems are currently being surpassed by RS6K, they aren't that far behind, and that's pretty good considering that they are still running at 700 MHz.
You can expect Itanium, with its 64-bit addressing, faster FSB, and support for massive amounts of SDRAM, to surpass Pentium III Xeon and take the lead, at least until 4-way Foster systems (and after that, 4-way McKinley systems) are released. Yeah I know, none of this has to do with the EPIC instruction set, but considering how late Itanium is, I think the potential TPC-C results should be quite an accomplishment in itself.
Tenchusatsu |
