[QUOTE=jinydu]Just wondering:

Why are personal computer speeds measured in Hz, while supercomputer speeds are measured in FLOPS? Why the diffferent units?[/QUOTE]

Because supercomputers usually have a lot of CPU while pc has 1. And when
you have thousands of cpus, it is not uncommon one or two of them arent
functioning.

[QUOTE=optim]ALU is Areithmetic-Logic Unit inside the CPU, it processes integer mathematical operations and boolean logic. MIPS is million instuctions per second. An instuction may be the addition of two numbers or a logical operation (AND, OR, XOR etc).

FPU is a unit inside the CPU for floating point arithmetic, that is, non-integer numbers. It measures in FLOPS.

Some CPUs, including Intel Pentium 4 and AMD Athlon64, have SSE2 capability. The SSE2 capabilities of a CPU can also be measured in FLOPS, because they handle floating point arithmetic at higher speed than FPU, but they are highly specific (i.e. FPU is general-purpose, SSE2 is for special operations).

So your CPU can compute up to 8161 MIPS when used for integer arithmetic and boolean logic, 2250 MFLOPS when used for general purpose floating point arithmetic, and 4110 MFLOPS when used for SSE2 arithmetic.

Also note than SSE2 accepts 128-bit instuctions, while ALU and FPU are only 32bit, If I recall correctly.

Prime95/Mprime can utilise SSE2 when available. However, it is faster in Intel P4 than AMD Athlon64. AthlonXP doesnt have SSE2.[/QUOTE]

So the 4.110 Gigaflops reading is the most appropriate one to compare with Primenet's 15.145 Teraflops?

[QUOTE=jinydu]So if FLOPS is a more useful measure, why not use it for PCs too? [/QUOTE]
FLOPS means less for a PC than for a supercomputer.
A supercomputer's primary purpose is to crunch (floating point) numbers, so the number of floating point operations it can do per second is of interest.
PC's are used for lots of things, and numbercrunching is typically just a small part of it. A typical PC program does a little computing, reads or writes a little (to disk or database), and does a little screen updating. Hence, FLOPS says less about 'how quick will my program run', since the variation in what PC programs spend their time on are so much greater.
Basically, FLOPS is not important for general PC users - common programs like Word and Internet Explorer do not use floating point operations at all.

But the emphasis on GHz instead is probably mostly the effect of Intel marketing strategy.

A weird thing about my computer is it seems to gain/lose more time than the speed differences would suggest. My computer seems to be running about even when it runs at 80 ms per iteration. But when I get it to run at 76.1 ms per iteratioin for 20 hours, my expected completion date goes down by 4 to 5 hours. That's 24 to 25 hours of work in 20 hours, so a 20 to 25% improvement. I should only expect about a 15 minute gain. Kinda interesting...

[QUOTE=jinydu]So the 4.110 Gigaflops reading is the most appropriate one to compare with Primenet's 15.145 Teraflops?[/QUOTE]
Yes.

The reduced completion time ( 4 or 5 hours ) is the amount of time saved
over the full time left for the test ( weeks/months ) using the slightly faster
time per iteration, not the gain for that one day.

A time savings over long period not per day.

[QUOTE=dsouza123]The reduced completion time ( 4 or 5 hours ) is the amount of time saved
over the full time left for the test ( weeks/months ) using the slightly faster
time per iteration, not the gain for that one day.

A time savings over long period not per day.[/QUOTE]

I don't think so. The remaining time has not gone back up, and it doesn't fluctuate during slower days at all. It seems to behave more like if it assumes you will be running the program at some constant speed, which on my computer seems to be about 78-80 ms per iteration.