Is it time to change the CPU year measurement?

[E_tron]

The Pentium 90mhz processor is long gone now, however we still use this processor as our standard unit of measure. Now we have teams that achieve well over 1000 P90 years. Is it time to use a new measure?

[jinydu]

Yes, I think we should change to a new measure. However, if we just change it to a new processor (say 1 GHz), that too will also become obselete eventually.

Thus, I would recommend changing the unit to PetaFLOPs (the last "s" is intended to signify a plural, not as an abbreviation for "second"). This seems to be the most logical choice, because it directly measures the number of floating point operations done. It also has the advantage that it wouldn't require too much of a change compared to our current system. According to my graphing calculator, 1 P90 year ~= 1.04*10^15 FLOPs = 1.04 PetaFLOPs. When dealing with larger quantities (such as the amount of computation done by TPR, curtisc, Primenet, or GIMPS as a whole), ExaFLOPs could be used, where 1 ExaFLOP = 10^18 FLOPs = 1000 PetaFLOPs. According to my calculations, as of May 26, 2004, Primenet has completed just over 500 ExaFLOPs. Maybe next year, we could celebrate reaching 1 ZettaFLOP (1000 ExaFLOPs).

It may seem a little strange at first, but people would eventually get used to it. Besides, it would make people more familiar with the less well known SI prefixes.

[xilman]

Originally Posted by E_tron

The Pentium 90mhz processor is long gone now, however we still use this processor as our standard unit of measure. Now we have teams that achieve well over 1000 P90 years. Is it time to use a new measure?

Every time this question comes up, the consensus seems to be that better the unit we know than anything else. It may be wrong, but at least it is definitively wrong, to misquote the Hitchhiker's Guide to the Galaxy.

The big advantage of the P90 year is historical continuity and, as pointed out, relatively simple conversion to other units. Jinydu has done everyone a service by publicising that a P90 year is one petaFLOP to within measurement errors, as this is particularly easy to remember. Leave it at that.


Paul

[E_tron]

Originally Posted by xilman

...The big advantage of the P90 year is historical continuity...

I actually think we should keep the P90. I mainly wanted to know how others felt about it. I think we should keep the P90, because there are many "enhancements" to modern CPUs. It would be unfair to measure in Pentium 4 years, because the P4 employs a net burst architecture(intel abandoned the 13 stage architecture). If we did convert to a new measure, the only fair measurement would be the Pentium 3@1.4ghz and AMD Athlon XP 3200+@2.2ghz, because they are closer to a traditional x86 chip.

[lycorn]

This has already been discussed in the foruns a while ago.(http://www.mersenneforum.org/showthread.php?t=16).
As you see, it is far from consensual. I personally vote for keeping our historical unit of P90 years, as any other would be as arbitrary, and would also get quickly outdated.

[ColdFury]

The whole P90 year measurement is pretty much arbitrary. Changing the measurement to some other CPU speed would be equally arbitrary, so I see no advantage of going thru the trouble to change it.

[jinydu]

Originally Posted by ColdFury

The whole P90 year measurement is pretty much arbitrary. Changing the measurement to some other CPU speed would be equally arbitrary, so I see no advantage of going thru the trouble to change it.

Yes, that's why I proposed changing to a unit that is independent of a particular CPU speed.

[ET_]

Originally Posted by jinydu

Yes, that's why I proposed changing to a unit that is independent of a particular CPU speed.

The number of floating point operatins is somewhat different, and dependent to the implementation of the microcode into the CPU. A vectorial processor, or even the APE 1000 would run Prime95 at maximum speed if correctly coded, but again you would compare tomatoes with oranges (I was writing apples, but it's not fair).

IIRC Prime95 was born to be used with Intel 486 and I think it is better to stay with the measure that gave the maximum spread to it, even if Alphas work on 64 bit and Crays use a different implementation...

Just my 0,02 €

Luigi