Awesome display of computing power!

Was looking at the team rankings and have to say that the world of computing has really had an effect on folding. I've never had a farm (some incredible setups around here) but have been folding for awhile. I remember when 1M points as a big deal. Now I'm inching into the top 200 and it requires about 2M points! To get into the top 100 requires about 4.5M!!

Awesome display of commitment folks!

The computing power is impressive, but it's really the points inflation that is awesome. I'm not over my first million points being devalued, by -bigadv and the quick return bonus, to less than 40 days work on an i7. If that's a true indication of the scientific value of those early points, then it wasn't worth it.

Yup ChasR, I think my 1st 1M took near 4 years and with the help from an eventual dual core and a 9800GT.

augie1111 said:

Yup ChasR, I think my 1st 1M took near 4 years and with the help from an eventual dual core and a 9800GT.

Click to expand...

I remember when it took a PS3, a couple of Celly D's, an AXP+, and a P3 just to make right in the neighborhood of 1300-1400 PPD. I was one of the late adopters of multi core.

I made 1 million my first year. At the end of that year, I had well over 100 instances of FAH running mostly on P4s and socket As. I think I made 2 or three million points with 100+ instances in the 2nd year. I figure I can completely sit out the i7 and 4xx upgrade cycle, piddle around with my old hardware and blow by everybody with new hardware and an exponential bonus points scheme in a couple of years.

Sorry but I think this is part of a much broader problem that isn't limited to folding points. If you guys know how to correct for inflation, please let me know. I think I've dollars setting around in a few locations that I would like to correct for inflation.

An i7 is about 150x more powerful than a PIII in terms of MIPS. When I first started I was making about 2,000 ppd with 20 celeron machines. Now the machines make about 100x that much per day, but with i7 and quads. Points seems pretty much in proportion actually.

The i7 is nowhere near 150 x as fast as a p3. In fact, their floating point performance (critical to FAH) is quite similar to a P3 in terms of whetstones per GHz/physical core. I researched this rather rapidly so somebody ought to check it out. Using a p3 since that was referenced, a p3 @ 1.0 does 21 Whetstones while a i7 @ 3.8 does 319. That would be 15 x faster. The p3 would be awarded 45 ppd while the 15 x faster i7 is awarded 26,000 ppd, 600 times more ppd! So even if we don't reconcile our decimal place difference in the relative performance, the evidence of massive points inflation is still clear.

I researched this a little more and found SiSoft results on various old and new cpus. Using sandra's whetstone numbers,

P3 1.0 GHz 1326
P4c Northwood, the "benchmark machine" 6252
17 920 stock 55530
i7 @ 3.9 95649

By this synthetic benchmark, the i7 is up to 72 times faster than a p3 and 15 times faster than a p4c @ 2.8 GHz. These numbers are a reflection of the relative raw floating point capabilities of the processors. The i7 SMP performance on -bigadv (27,000 ppd) awards 600 and 245 times the ppd respectively.

ChasR said:

I researched this a little more and found SiSoft results on various old and new cpus. Using sandra's whetstone numbers,

P3 1.0 GHz 1326
P4c Northwood, the "benchmark machine" 6252
17 920 stock 55530
i7 @ 3.9 95649

By this synthetic benchmark, the i7 is up to 72 times faster than a p3 and 15 times faster than a p4c @ 2.8 GHz. These numbers are a reflection of the relative raw floating point capabilities of the processors. The i7 SMP performance on -bigadv (27,000 ppd) awards 600 and 245 times the ppd respectively.

Click to expand...

The bad thing about all of this is that I didn't even see 5 digits for points until after I started folding on the PS3. When all of that got shut down, and I moved to the Q6600 and 9800GT, I went from barely seeing anything for PPD, to around 10k. Now, I'm not even sure what the PPD on my Q6600 would look like, since I can't run 24/7, so it would just be a waste of WU's to fire up the SMP client.

ChasR said:

I researched this a little more and found SiSoft results on various old and new cpus. Using sandra's whetstone numbers,

P3 1.0 GHz 1326
P4c Northwood, the "benchmark machine" 6252
17 920 stock 55530
i7 @ 3.9 95649

By this synthetic benchmark, the i7 is up to 72 times faster than a p3 and 15 times faster than a p4c @ 2.8 GHz. These numbers are a reflection of the relative raw floating point capabilities of the processors. The i7 SMP performance on -bigadv (27,000 ppd) awards 600 and 245 times the ppd respectively.

Click to expand...

I was just guessing about 150x but according to wiki, the p3 500 mhz is 1,354 MIPS and i7 @ 3,300 mhz is 147,600; so that's about 110x faster. In addition, things like larger cache, faster memory and more efficient memory transport add more speed to memory hungry projects than the MIPS account for. Totaled the speed increase is probablly comparable to 150x or more.

In addition, the i7 -bigadv points are extra bonus points, which won't last forever and don't apply to many folders - so don't apply general inflation to everyone. Also, the new WUs like p6701's take you down to around 12,000 ppd which significantly decreases the top multiple to around 250x to 100x - which is comparable to the actual system speed increases with the i7.

THe total instructions per second means very little to FAH. Floating point instructions are what is critical.
If you run uniprocessor clients on the i7, you really do get the performance difference between the i7 and the p3 or the p4c. You won't get 12,000 ppd, you'll be making more like 3000 ppd. 12,000 from the smp client vs 3000 from the uniprocessor client is indicative of significant inflation of the points awarded for the work done.

I think the points inflation comes into play because they want to encourage us to run the newer projects. Newer projects require more processing power to finish by the deadline so they set it up so we can collect a bonus if our hardware is fast enough.

ChasR said:

12,000 from the smp client vs 3000 from the uniprocessor client is indicative of significant inflation of the points awarded for the work done.

Click to expand...

That assumes that speed has a linear relationship to value. But, the value of speed has a premium, just as Intel charges $1k for a fast processor and only half that price for a CPU only slightly slower. Speed costs $ and you should be rightly rewarded for adding speed to the Pandi project - assuming speed is valuable to them. Pandi has often said speed is the most important factor to them - or example, asking contributors not to run two UNI clients on a hyperthreaded machine because the results come back slower. The new bonus system totally rewards speed, so it must be very important.

Pandi Group decides how valuable each type of work and its return speed is to them, so its not possible for us to determine the degree of point inflation vs better/more science being done faster. Presumably -bigadv science is much more valuable to science than what a UNI produces.

You've actually hit on the principal error of the points scheme and the primary reason I'm simewhat ranting about points inflation. Bigadv wasn't intended to be run on desktop processors. It was intended to run on server class multi socket machines and the points were set to encourage a somewhat limited number of donors to run it. When designed Pande Group didn't know Bigadv would run on the i7 much less how well.
Anyway the points inflation is real and it is demonstrably a large number. When SMP was introduced, I was making about 3600 ppd on my best Q6600s, now I'm making 7500 ppd on the same machine, kind of hard to argue with that. Pande group has used inflation repeatedly to influence donor behavior to further the science (it works on me when I have money). We'll likely undergo another round of inflation when the QRB is introduced to uniprocessor clients. If the trend continues as it has since FAH's beginnings, the points we produce today will be worth less than a quarter of the points 3 years from now, such is the nature of exponential bonuses. Now if they'd just go back and apply the QRB to all my old WUs, if shutup in a heart beat.

I managed to climb to 685 world wide, but now I've been stepped over by all the newest i7 and GPU3 folders and have dropped below 750 and I'm still sinking.

My 1xQ6600 and 9800GTX+ are only good for 10k-11kPPD and not competitive anymore.

What I used to get PPD a few years ago is a joke today.

How much further is this point inflation going to go?

Exponential curve perhaps.