My PS3 has folded 293 work units. What about you?

Surely there can't even be any doubt. The average throughput per GPU on F@H is far higher than the PS3 and since the average GPU contributing will be weaker than an X1900XT the conclusion seems obvious.

I don't know how CPU's compare at the high end but a properly scaling client should be able to fold 2-3x faster on a Q6600 than a fully utilised X2 4400+ (assuming it is utilising both cores.)

I said upgrade. Q6600 = $285, X1950Pro = $124 The Pro isn't quite an XT but for folding it should be pretty close. I guess you might need a new mobo and RAM too if you have an older PC.

 

Well according to Sandra my box does 33556 MFlops (iSSE3) and 47189 MIPs (Dear God! I remember when the 68040 came out and did like 4 MIPs, I nearly wet myself then... So much power!!) .

Anyway that's 8389 MFlops per core (@2.6GHz) (very roughly speaking). The CELL was rated at about 218 GFlops with 8 SPU's and at 3.2Ghz. So if we cut that back to about 190GFlops it's still far in excess of the 8.5 GFlops per core my PC can do.

So in terms of computational power the CELL is a monster. It's just the type of WU's that it can do that will be the difference. CPU - Highly Sequencial. GPU - Highly Parallised. CELL - somewhere in the middle?

 

Let's put it this way:
my measly P4 HT 3.4GHz Prescott pumps out a P2563 SMP WU (1760 points) in about 3 days, with me using the system several hours daily, often not folding at all whilst MP gaming. How many points does a PS3 put out in the same 3 day period? Now compare that to an overclocked quad-core system putting out 12,000 points in the same amount of time.

Yes, an overclocked quad (3GHz+) will put out 4k PPD or more when running SMP folding. I'd like to see how many PS3s it takes to match that, and I guarantee the number isn't 1

As for GPU folding: at this point in time it isn't worth it. SMP folding (even on a dual-core) returns far superior yields. Even Pentium Ds can pull off 1k PPD. For a point of reference, my X1650 XT (lowest grade GPU which will fold) achieves ~11min frame times w/GPU folding, yielding ~ 400PPD. I get at least 500PPD SMP folding on my P4, up to almost 600 PPD. The fastest folding GPU (X1950 XTX) gets only 660 PPD at stock speeds (completes 1 GPU WU worth 330 points in 12 hours). Not chump change, but certainly no match for SMP folding on a capable CPU (which my P4 clearly is not).

 

Cell is a bit more general-purpose than GPUs are for folding, with traditional CPUs taking the crown in terms of variety of code they can execute. Cell is certainly a computational monster, but its low DP rate (compared to its SP rate) means it gets out-folded by traditional dual-core CPUs (when running SMP folding), and blown away by o/c'd quads (see previous post).

 

Is there any way, other than direct observation of your stats on a daily basis, to figure point per day? I've got a Q6600 that I can run 4x F@H command line (linux) to compare with the PS3.

Either way for the $$ and energy (with the new 40 GB model) the PS3 is a darn nice way to get more WUs done.

 

Stanford stats page. If you're ranked high enough, you can check your stats at EOC which has nice fancy graphs and a whole lot more information.

More WUs does not equal more work. Points are the best way to look at work performed. Also, you really should try running SMP folding. You'll get way more points than you will just running multiple instances of the standard client. As yet another point of reference for my P4 rig, standard 206 point WUs (the WU du jour of late) yield ~ 110 PPD for me. I'll take 5x that amount via SMP folding, thank you very much.

 

Is there a linux smp client?

Already know about the stats page - that's what I meant when I said direct observation - i.e. delta by day is manual.

 

Not only is there a Linux SMP client, it's actually the preferred client. Download

F@H SMP uses the MPI model, since Linux natively supports this and Windows has to rely purely on F@H for MPI support, Linux SMP tends to be faster and more stable.

Check EOC's stats page. As I said, much more info than Stanford's. Weekly totals (points and WUs), monthly totals (points and WUs), daily averages (points), and a graph plotting your PPD as well as a threat/conquest list.

 

EOC's is nice. The SMP client wants 64-bit linux...shame as I'm running 32 bit for app compatibility.

 

Could always run it in a VM as long as you have a 64-bit proc, you can host a 64-bit Guest OS on a 32-bit OS through VMWare Server (free). As long as you've got the system RAM, of course (1GB per host/guest OS is recommended, although you can get by with 512MB per if needed).

 

I've got the memory (4 GB), but when I upgraded 7.04->7.10 (ubuntu) it borked my vmware server install somehow and now 7.10 doesn't have an ubuntu specific vmware server distro so I'll have to uninstall vmware and install 1.04 from the vmware site. pita.

 

So tempted to switch over to a 64bit flavour of linux, any suggestions at to which distro would be good to go as a dual boot option for Vista (teh suckiest os ever).

 

The flavor of the day seems to be Ubuntu.

 

657 WU's now.

 

I've got 59 on my PS3 as of last night, folding basically 24x7 (obviously not while gaming though) for <1 month. It's nice, but Stanford needs to re-work their point system as PS3 WUs don't earn nearly enough points. It pisses me off that there are people out there running SMP F@H on their quads and getting well over 4k PPD, when my 30+ CPU farm (including the PS3) all folding and contributing WUs on a weekly basis can barely pull of 3k PPD. Bullshit.

 

700 today!

 

read the stuff that's been said...

Doesn't that sound a bit contradictory? You said yourself that points were the best judge of work performed. Stanford's reason for making the PS3 workunits worth less is simply because the information found through it is not as valuable. The cell cannot handle the giant spectrum of calculations that desktop cpu's can, thus there are a lot more things Stanford can do with SMP folding workunits.

Why does it piss you off that people have put money into buying quad core systems and are choosing to devote their cpu's time to scientific research with no monetary reward?
Also, why does it really matter how many points you get? The project wasn't designed to boost people's self image, but rather to find out more about various diseases that have caused millions of people to die...

 

I'm well aware that the ISA of Cell's SPEs can't handle the variety of calculations that the average multi-core x86 MPU can, and that simulations on the PS3 are limited by the (relatively) small memory pool, but the sheer number of FLOPs alone and the recent PR boost PS3 folding has brought to the project justify a point boost in my mind.

You mis-understand. I'm not pissed that people are utilizing their multi-core x86s to run SMP folding, I'm pissed that Stanford has chosen to award these WUs such a large point bonus that long-time folders with many systems are bumped off the PPD leader board by these people. I can assure you that the 20+ systems I have folding and submitting WUs on a weekly or even daily basis are contributing more useful work to the project than a single quad core system can, yet the points system currently in place does not reflect this reality.

The point system is a means of drawing upon the competitive nature of people to get more work done by getting people to fold on as many systems as they can. Folding teams are a testament to this fact.

 

I just happened to dip into this forum seeing a Cell folding thread highlighted, and now I'm compelled to post outside my usual haunt!

The Cell can handle every bit of maths possible on a 'desktop CPU'. Every single bit. If they want IEEE standard double precision they can code it. The reason Cell isn't put to work on the same work units is because when optimized, certain calculations can run much faster. It's deemed more efficient to set the PS3's working at 10x the usual rate on a (fairly large as I understand it) subsection of total workload than run at 1x the usual rate on the same full range of problems as the standard CPU implementation. PS3s churning through the bulk workloads while PCs work on the tricky stuff is attaining much better overall progress than throwing everything at the much slower complex systems. But whatever choices Stanford are making, be sure that there are no maths limits on Cell. There's not a single mathematical operation that it cannot do. Indeed there's no processor out there with proper basic maths support that can't handle every mathematical operation, as all maths is built from the fundamental, atomic operators. It's a matter of performance in these maths that is the limiting factor, and not actually ability of a CPU to be able to perform mathematical functions.