<<
Unfortunately, all RISC chips(including SPARCs) have rather poor Integer units. >>
check out
Spec Int Scores
Compare these: (base, and peak)
Intel Corporation Intel OR840(1 GHz Pentium III proces) 1 438 442
Alpha Processor, Inc API UP2000 750 MHz 1 434 456
Dell Precision WorkStation 330 (1.50 GHz[P4] 1 515 526
Sun Microsystems Sun Blade 1000 Model 1900(900mhz) 1 427 482
Hewlett Packard Corp hp visualize j6000 UNIX workstation 1 417 441 (552mhz PA-RISC 8600)
Clock normalized (linearly....not technically fair, I know)
P3 @ 1ghz 438 442
Alpha 21264A @ 1ghz 521 547
P4 @ 1ghz 343 351
Ultrasparc III @ 1ghz 474
PA-RISC 8600 @ 1ghz 755 873
Wow, what do we have here...
The P4 is at the bottom of the pile on a clock normalized basis, but we knew that. Take a look at those scores....who's integer performance, per clock, is the strongest? Wow, HP's!
The point being, shear mhz can make up for IPC. The two are directly proportional. HP went with the high IPC, low clock rate (and MAMOTH die sizes I might add, because of
1.5mb of
L1 cache.) The P4 on the other hand, is the exact oppposite. And there's a reason for that. x86 code DOESN'T GENERATE VERY MUCH ILP generally speaking. So instead of going for high IPC, which is costly in terms of die size, and the fact that its just really hard to do, Intel went for high clock rate.
RISC chips aren't weak in integer performance! They are, infact, generally faster per clock. x86's and CISC chips in general, have been saved only by using BETTER process technologies which allowed them to keep the clock rate up.
The reason why RISC chips don't do so well is that, as far as I know, they don't generally have an equivalent of the ROL x86 instruction (why? Because its not used often, and therefore, the functionality of that instruction doesn't need to be fast, and can be done via other methods, albiet slower, which is alright because its not used often! AKA, the RISC principle!).
[/end rant]
Facebook
Twitter