wuttz

did a little experiment today on the fastest software i've seen to calculate pi values to 1m/32m digits.

first off, credit goes to amdzone.com(abinstein), where i learned about this.



i've not seen any windows pi calculator come close to this kind of performance. this is using the gmp-chudnovsky.c program from: http://gmplib.org/

my scores(from image above)
1m pi digits total time = ~1sec(HWBOT WR=6.75secs w/ LN2)
32m pi digits total time = ~72secs(HWBOT WR=375.30secs w/ LN2)

my settings are all at stock (DDRAM=1066, 7-7-7)with my sig rig.

09/05/2009
Now why do I say SuperPi is useless?

SuperPi is an antiquated benchmark that assesses core performance using an antiquated fpu x87 instruction set. It is a grossly inefficient computer program that uses floating point units to calculate pi and "evaluate" processor cache performance. It's not by any means a "real world" reflection of todays modern cpu's microarch capabilities. It's like running a horse(fpu x87 instructions) down today's paved interstate(the cpu hardware), when all the while a car(newer simd's) are available.

SuperPi fails to take advantage of newer instruction sets, and hence also fails to evaluate modern-day hardware capabilities that newer processors support and are designed with. The FPU X87 instruction set was developed for an external co-processor dedicated to computing mathematical tasks and worked with the 80286/80386 microprocessors. Now if you havent heard of the 286/386, then you know SuperPi is that old.

How many real world applications today still use the FPU x87 instruction set? Moreover, FPU x87 legacy support is still integrated in todays microprocessors since it uses less than 5% of the microprocessor die. The headache in legacy support is not worth the cost of removing it from the die, so it stays there. However, why dont we benchers use a better benchmark tool that uses at least more than 5% of the processor die(and updated code)? Or are we just too involved in saying, hey 5% of my processor runs really really well!



Lastly, with SuperPi being closed source you can't inspect the code for any vendor optimizations, so it is inherently suspect. To make matters worse, it is mostly compiled on vendor-biased compilers (which is all we have in a Windows world).

Now my point is, maybe we have focused too much on the hardware and continually do hardware upgrades that we forget the fact we are using a 15 year old outdated benchmark to evaluate said state-of-the-art equipment? Using SuperPi to "benchmark" processors is really just shoving garbage, age old, code down the latest processor that you have. Unless the typical computer user today still likes to use Windows 3.1, then I can understand the significance of SuperPi. But alas, 'tis time to move on. It's the Windows 7 era already.

Shouldn't state-of-the-art hardware go hand in hand with cutting edge benchmarking software?

Or is the pursuit of performance really just on the hardware side?

[Will be updated as I learn more about CPU architectures and software.]

__________________

aaronmonto

Is the Phenom II still 35% slower than C2D's at the same clock speed in this benchie?

__________________

wuttz

http://gmplib.org/pi-with-gmp.html

theres times listed there, although its not comprehensive. if we have intel benchers interested, i'd appreciate your input.

__________________

DuckieHo

A benchmark's purpose is NOT get the best score. It is to evaluate performance of different hardware. SuperPi testings floating point and cache performance. I don't believe it is multi-threaded. If you want even faster PI scores, there are certain methods that can be broken apart into threads. Those would yield faster times.

i.e. You don't benchmark a car's 1/4 mi time on downhill slope. Yeah, it would reduce the times for all cars but that's not the point.


There are a number of different ways to caculating PI.
The one used in that benchmarks is this: http://en.wikipedia.org/wiki/Chudnovsky_algorithm

Here's at least another dozen methods: http://en.wikipedia.org/wiki/Numeric...ions_of_%CF%80

__________________
To answer most of your questions: (1) a fridge cannot cool a PC (2) 64-bit OS for over 3.4GB (3) If a PCIe card fits, it should work (4) Resolution, not screen size (5) If you have a question, it is not news (6) Report, not respond to Spam (7) Single-Rail/Non-Modular PSUs are not always better than Multi-Rail/Modular

aaronmonto

You just explained the world to me... The Phenom II's L3 is slower than my RAM. No wonder.

__________________

wuttz

i believe the better example would be having a "horse" as the benchmark software(superpi), and running that horse down a paved interstate. is that "horse" really "evaluating the true performance of the paved interstate?"

also, gmp-chudnovsky.c is an open source program with the code open for inspection, just to be sure no vendor-optimized code gets dropped in.

on top of that, it is compiled on a vendor-neutral opensource compiler, gcc.

can you say the same for superpi?

__________________

wuttz

did a comparison run using y-cruncher, a windows pi calculator client that makes use of the same chudnovsky compute algorithm i used under linux.

first, a core-for-core comparison between the two operating systems.



the linux bench was faster by 5 seconds in computing 32m pi values.

now, turning on multi-threading for y-cruncher..



it finishes the task in ~21.3secs!

these programs clearly make use of more efficient code and up-to-date instruction sets that modern computer architectures have. so i'd say these pi calculators reflect a better assessment of a cpu's computing abilities, instead of superpi.

__________________

Starbuck5000

Like what Duckie said.

Competative benchmarking isn't really about a showing the most accurate reflection, since when have sythenic benches really be accurate to real world performance anyway? If that was the case then we would only be benching Vantage.

Really competative benchmarking is more about the tweaks and the skill in tuning the system. PI is a benchmark that responds well to this. SuperPI is a single threaded benchmark. If you wanna see it added to hwbot make this same post on their forums

__________________
Electrical Engineering and Computer Science Student (Joint Bsc Hons)
STUDENT FINANCE HURRY UP!!!

Has far to many components for a 20 year old, and while you're at it forgive my spelling, the keyboard is dyslexic

BenBrown

I tend to agree more with wuttz. Standards need to be updated every once in a while. This should definitely be posted on HWBot.rg.

__________________
"You make a living by what you get. You make a life by what you give." --Winston Churchill

3DMark Vantage - 19,552 -- 3DMark06 - 25,066 -- 3DMark05 - 31,688 -- 3DMark03 - 114,287 -- CPU-Z - 4800MHz

** Stuff For Sale **
**The Brown Blogger**
Phenom I vs. Phenom II - Clock for Clock & Phenom II Overclocking Results

CL3P20

I dont know what to think...I am getting .541sec on the 1mil Pi calc... ....more testing will need to be done on the nature of the tuning that most affects the outcome of the score [other than CPU speed].

__________________

..:Q6600 Club:..

INTEL 45

e8400: 5.13ghz
e7400: 4.5ghz
Q6600: 4.44ghz
8800GS@1028mhz
My photo albums