How well could X360 handle F@H?

[patsu]

Just curious, how many work units can a GPU client churn out per day ?

Answering my own question...

http://techreport.com/etc/2006q4/gpu-folding/index.x?pg=1

These guys completed 8 work units (2640 points) in roughly 5 days on an X1900XTX. I just submitted my 8th work units in about (less than) 3.5 days -- music playing and gaming time included. My current score should be around 2400+ points.

Not sure if they are comparable though (Is the PS3 client really based on the GPU one) ? Upcoming GPUs should be faster too.

 

[V-G]

Curiously, the CELL code is actually derived from the GPU code (one of the reasons they are limited to the same types of WU's).

Funny that I have to tell someone like Dave this; please when making these kinds of statements you need to back them up. Appreciated.

Would also be nice with some reasoning, a guess would be time constraints.

 

["Nerve-Damage"]

Honestly, as much as I love folding on both my PS3s, I somewhat wish Sony didn’t get mixed up in the whole folding scene (console wise anyway). Its potential stuff like this (Versus vs. Versus stuff, Fan*** vs. Fan***) that will detract from the real purpose of folding. I can see it now; one side claiming the other side is fudging/cheating and the universities, institutions, etc…are in cahoots with them.

 

[Karma Police]

Honestly, as much as I love folding on both my PS3s, I somewhat wish Sony didn’t get mixed up in the whole folding scene (console wise anyway). Its potential stuff like this (Versus vs. Versus stuff, Fan*** vs. Fan***) that will detract from the real purpose of folding. I can see it know; one side claiming the other side is fudging/cheating and the universities, institutions, etc…are in cahoots with them.

I think that the opposite seems to be happening. I've seen more guys that hated Sony starting to like the PS3 when the total TFLOPS were being released. Could F@H be the app that brings PS3 & X360 haters together in perfect harmony? But even the worst fa**oyism is worth getting 30 million X360's & 30 million PS3's folding, anyway.

 

[patsu]

Honestly, as much as I love folding on both my PS3s, I somewhat wish Sony didn’t get mixed up in the whole folding scene (console wise anyway). Its potential stuff like this (Versus vs. Versus stuff, Fan*** vs. Fan***) that will detract from the real purpose of folding. I can see it know; one side claiming the other side is fudging/cheating and the universities, institutions, etc…are in cahoots with them.

True. People should be able to discuss this without bringing in hate. I am interested to know the effectiveness of the Cell architecture in general. I also fully expect the GPUs to overtake Cell in specific tasks (beyond graphics) due to the faster refresh cycle and advance in technology.

 

[V-G]

True. People should be able to discuss this without bringing in hate. I am interested to know the effectiveness of the Cell architecture in general. I also fully expect the GPUs to overtake Cell in specific tasks (beyond graphics) due to the faster refresh cycle and advance in technology.

But then again. has there actually been any hate?

 

[mc6809e]

but GPUs only do a subset of all possible folding calculations

That's just it, though. According to quiet a few people, ATi's GPUs are several times more powerful than the Cell at protein folding:

http://www.dailytech.com/article.aspx?newsid=6609

Whoa, there. From the F@H FAQ:

...they achieve this high performance by losing generality -- there are only certain types of calculations which would be well-suited to GPUs.

So sure, GPUs are faster -- if you don't mind ignoring the the more challenging computations that Cell can handle.

Similarly limiting Cell to the easy calculations might provide a similar speed up.

 

[Dave Baumann]

Not sure if they are comparable though (Is the PS3 client really based on the GPU one) ? Upcoming GPUs should be faster too.

Irrespective of what the client is based on or the type of cores that can be processed, it doesn't mean that the WU's are the same. I've not seen if any PS3 WU's crossover with GPU WU's. You can get different WU lengths within the same type of core as well (there's less variability with GPU/PS3 clients as there are fewer core types than those executed on standard CPU's).

Scoring is not comparable either; the scoring for the "high performance clients" isn't based on throughput relative to the traditional CPU clients, but much more arbitary criteria (unfortunatly) and I find the other day PS3 scores higher than GPU's because its a "more stable platform" - although now this is ring fenced in our drivers and there are other things we are looking at that will be solve this and considerably increase performance at the same time by removing the DX API layer. If something were to be done on the 360 then it probably wouldn't be as great an overhead as the API layer is much thinner.

Finally, the Tech Report article you link to doesn't mention any tweak application installed, so the performances are probably run at the clocked down 2D speeds.

Funny that I have to tell someone like Dave this; please when making these kinds of statements you need to back them up. Appreciated.

I've been told it by those involved. Sony did much of the work for the PS3 port, but the code that was supplied orginated from the GPU code, not CPU.

 

[digitalwanderer]

although now this is ring fenced in our drivers and there are other things we are looking at that will be solve this and considerably increase performance at the same time by removing the DX API layer.

Sorry if this is another stupid of mine, but what do you mean by "ring fenced"?

 

[patsu]

Irrespective of what the client is based on or the type of cores that can be processed, it doesn't mean that the WU's are the same. I've not seen if any PS3 WU's crossover with GPU WU's. You can get different WU lengths within the same type of core as well (there's less variability with GPU/PS3 clients as there are fewer core types than those executed on standard CPU's).

IMHO, WU completion duration is probably a better way to indicate efficiency and effectiveness (rather than FLOP counting). If the source is based on GPU, what separates GPU WUs from PS3 WUs ?

 

[patsu]

Curiously, the CELL code is actually derived from the GPU code (one of the reasons they are limited to the same types of WU's).

Some official statements about Cell's generality
http://forum.beyond3d.com/showpost.php?p=955700&postcount=256

 

[Dave Baumann]

Sorry if this is another stupid of mine, but what do you mean by "ring fenced"?

Making sure that the driver is in a state that F@H likes if its running.

IMHO, WU completion duration is probably a better way to indicate efficiency and effectiveness (rather than FLOP counting). If the source is based on GPU, what separates GPU WUs from PS3 WUs ?

Projects. As I said, I don't know if any PS3 projects cross over to GPU projects hence whether they process the same WU's. As your lik points to, the GPU client is still the fastest client, but that doesn't mean the scores are going to reflect that.

 

[patsu]

Projects. As I said, I don't know if any PS3 projects cross over to GPU projects hence whether they process the same WU's. As your [link] points to, the GPU client is still the fastest client, but that doesn't mean the scores are going to reflect that.

Yes... I suspect so -- as indicated in my previous post in this thread, but I was (still am) trying to find out how much faster.

The FLOP counts don't say much if the F@H folks use subjective instruction counting, elastic time (People who interrupted the calculations will get lower FLOP count), and implement different algorithms due to architectural differences.

The link also confirmed that the Cell CPU is indeed more general.

 

[Dave Baumann]

The link also confirmed that the Cell CPU is indeed more general.

The primary issue that the the "High performance clients" is that of effective parallelism - although there is the SMP client, the main way that PC users get higher throuput from multi-threaded/multi-core CPU's is just purely by running same many clients as there are physical/virtual CPU's - which is likely why the Cell version was derived from the GPU.

 

[patsu]

The primary issue that the the "High performance clients" is that of effective parallelism - although there is the SMP client, the main way that PC users get higher throuput from multi-threaded/multi-core CPU's is just purely by running same many clients as there are physical/virtual CPU's - which is likely why the Cell version was derived from the GPU.

Yes... for the implicit solvers, they took it from the GPUs. Cell will also have explicit solvers from the CPU side (or from scratch) according to the link.

 

[zed]

IMHO, WU completion duration is probably a better way to indicate efficiency and effectiveness (rather than FLOP counting). If the source is based on GPU, what separates GPU WUs from PS3 WUs ?

yes FLOPS aint a true gauge of the actual amount of work done, one of the stanford guys mentioned with the GPUs they recalculate the same equations multiple times (due to lack of being able to write anywhere in memory) with a CPU u just do the equation once, im not sure what method the ps3 uses.
its a pity that all 3 implementations dont use the same data

 

[bkilian]

Using memexport, you could probably get around the problems that GPUs don't do so well by doing a hybrid GPU/CPU solution. We do that to pretty good effect in the HD DVD player on the XBox. We're doing a fair amount of the VC1 and H.264 decoding tasks on the GPU while doing the rest (like the CABAC) on a few of the hardware threads. Add in a full XML/Javascript engine, a second video stream, and audio decoding, mixing and re-encoding, and even we're sometimes amazed it all works.

 

[patsu]

http://forum.beyond3d.com/showpost.php?p=955041&postcount=17

Yes, I think so too. The F@H does not require intensive graphics, so the implementor should be able to use Xenos more for calculations.

 

[Karma Police]

After talking to Mike at Stanford F@H, he said that they have contacted MS a few times to setup an X360 client, but that MS has not shown any interest. However, as of Oct 06 someone at MS has contacted Rosetta@Home about setting something up there. But I think there hasn't been any activity since.

Even if MS set it up as a "game" on arcade and charged $5.00 I'd pay it.

bkilian, do you know if MS has any plans for this? All of the DC communities would really like the help. What would hold MS back?

 

[SirTendeth]

Here are two links relating to GPGU functionality and Xenos.

http://www.doggetts.org/michael/unc-xenos-doggett.pdf

http://www.gpgpu.org/sc2006/slides/09.owens.tips_and_tricks.pdf

In the first link it mentions doing vertex scatter writes (probably MEMEXPORT but doesn't specify), and in the second, well the second would take someone more versed in GPGU funtion to know how Xenos may be better utilised in Folding.