X1800XT up to 7X faster than 7800GTX

Printable View

10-07-2005, 01:58 PM
Ugly n Grey

1 Attachment(s)

X1800XT up to 7X faster than 7800GTX

Distributed computing programs are nothing new and neither are discussions about using the VPU of your video card to fold. A recently leaked document from the Stanford University Graphics Lab has the X1800XT leading it's rival the GTX by over 7x in certain types of calculations with performance equivelant to a 10.5 GHz P4 CPU.

Maybe the overall architecture lends itself better to scientific calculations or maybe the X1800 series is just everything ATI claims. In any case it looks like it might be THE graphics card for folding at home applications in terms of performance. This could mean a huge boost for Folding at Home teams all the world just like OURS here at XS that recently passed over it's one millionth work packet.

See the attached PDF for details.

http://img230.imageshack.us/img230/9...acsperf4ps.png
10-07-2005, 02:01 PM
Vapor

That's pretty freakin sweet--too bad D2OL isn't for GPUs yet, I'd definitely pick up an XT at a little over retail just to help the cause (and my benchmark scores, of course ;)).
10-07-2005, 02:04 PM
Cooper

Impressive.
I wonder how much of that power is used in todays games if XT beats GTX for ~15-20% ???
10-07-2005, 03:14 PM
P_1

Quote:

Originally Posted by Ugly n Grey

Distributed computing programs are nothing new and neither are discussions about using the VPU of your video card to fold. A recently leaked document from the Stanford University Graphics Lab has the X1800XT leading it's rival the GTX by over 7x in certain types of calculations with performance equivelant to a 10.5 GHz P4 CPU.

Maybe the overall architecture lends itself better to scientific calculations or maybe the X1800 series is just everything ATI claims. In any case it looks like it might be THE graphics card for folding at home applications in terms of performance. This could mean a huge boost for Folding at Home teams all the world just like OURS here at XS that recently passed over it's one millionth work packet.

See the attached PDF for details.

I wonder how an Athlon 64 stacks up.
10-07-2005, 03:40 PM
JuanFlaiter

I thought the X1800XT would be 7x faster than 7800GTX, when dropped from a tenth floor. xD
10-07-2005, 03:44 PM
twiggy

:eek:

Next we'll see folding farms with x1800xt's in X-fire! :slobber: :hehe:
10-07-2005, 05:02 PM
freecableguy

:lol2:

Quote:

Originally Posted by JuanFlaiter

I thought the X1800XT would be 7x faster than 7800GTX, when dropped from a tenth floor. xD

Hahahahahahhahaha.,........ :clap:
10-07-2005, 05:27 PM
One_Hertz

Quote:

Originally Posted by Cooper

Impressive.
I wonder how much of that power is used in todays games if XT beats GTX for ~15-20% ???

More like 10 on average... It just shows how inefficient the architecture is maybe? :confused:
10-07-2005, 05:28 PM
Welz

wait, so we can vold with our gpu's? lol i think i missed something here
10-07-2005, 05:44 PM
Ugly n Grey

Quote:

Originally Posted by One_Hertz

More like 10 on average... It just shows how inefficient the architecture is maybe? :confused:

Maybe how innefficiently we use graphics and drivers.....there's a lot going on when you have to format output for a screen, when you don't, you can get away with murder....

As far as the dropping comment, it was funny :D, but didn't make a lot of sense in this context really. More of a thread crap, please see the Xtreme Toilet for your needs.

To the question on how an AMD 64 stacks up, on this type of calculation it's pretty close to a RISC chip 1:1 so pretty much tied with a G5 at the same frequency for this type of thing, maybe within 90% of the G5. Not for general use of course (A64 is better) but for science PPC is hard to beat really. One of the reasons for that is they are usually running Unix 64 bit, so it's more about the architecture and the use than the actual chip.
10-07-2005, 09:30 PM
Mikesta

Quote:

Originally Posted by P_1

I wonder how an Athlon 64 stacks up.

Athlon 64's stack up badly against P4's in F@H. This is because F@H uses Intel compiler that does not allow SSE, SSE2 & 3 to be used with non-intel processors.

So A64's have been shafted VERY badly. Even when competing on certain Gromacs where SSE is not that big an issue P4 still pulls ahead around 10-15% against equivalent a64's. (i.e. P4 3.2 vs a645 3200+).

F@H also loves cache so A64's are weaker compared to Intel proc's.

Stanford better get their asses in gear and get something that helps us AMD users produce more.
10-07-2005, 09:49 PM
P_1

Quote:

Originally Posted by Mikesta

Athlon 64's stack up badly against P4's in F@H. This is because F@H uses Intel compiler that does not allow SSE, SSE2 & 3 to be used with non-intel processors.

So A64's have been shafted VERY badly. Even when competing on certain Gromacs where SSE is not that big an issue P4 still pulls ahead around 10-15% against equivalent a64's. (i.e. P4 3.2 vs a645 3200+).

F@H also loves cache so A64's are weaker compared to Intel proc's.

Stanford better get their asses in gear and get something that helps us AMD users produce more.

I see, but aren't they making a custom version to run on the x1800xt? If they can do that, why not make a custom AMD version?
10-07-2005, 10:12 PM
STEvil

Quote:

Originally Posted by Mikesta

Athlon 64's stack up badly against P4's in F@H. This is because F@H uses Intel compiler that does not allow SSE, SSE2 & 3 to be used with non-intel processors.

So A64's have been shafted VERY badly. Even when competing on certain Gromacs where SSE is not that big an issue P4 still pulls ahead around 10-15% against equivalent a64's. (i.e. P4 3.2 vs a645 3200+).

F@H also loves cache so A64's are weaker compared to Intel proc's.

Stanford better get their asses in gear and get something that helps us AMD users produce more.

would this be a good application to test the geniune intel check remover then (bypasses this check which allows intel cpus to use sse/2/3 but not other cpus)?
10-07-2005, 10:20 PM
agenda2005

DANG DANG DANG!!! GOOD FIND. WAO I didn't know that these X1800's will pull this kind of stunt! I may have to pick up two X1800XT and pair them with a dual core-dual CPU(4 CPU) Opteron. Offloading all the force calculations to the GPU will definitely do some damage in molecular dynamics. Thumbs up to the Stanford team. I've been following them, but previous results were not as encouraging as this new ones. ATI get my money this time.
10-07-2005, 11:04 PM
Lithan

That's one of the wierdest things I've seen lately. Might be some insight into the architecture of ati's chips.

One hertz, cooper, where are you getting these numbers from? I haven't seen any benches yet and really want to see some.
10-08-2005, 01:51 AM
Mikesta

Quote:

Originally Posted by STEvil

would this be a good application to test the geniune intel check remover then (bypasses this check which allows intel cpus to use sse/2/3 but not other cpus)?

HELL YES!

Plz do test it. It would help me, and others immensely!!!!
10-08-2005, 02:37 AM
MaxxxRacer

pretty freeking intersting....

Just wish the benches relfected this.. but if they did, they could charge 900 dollars for an X1800XT and still sell them faster t han they could build them.
10-08-2005, 02:43 AM
Cooper

I agree with UnG on inefficient hardware use.
Lithan, actually I pulled this number out of my finger - I`m really confused with all those reviews. IMO non of todays alpha games have the ability to fully use R520 hardware.
But I think with the right software/drivers X1800 could be more than 20% efficient against GTX. It`s just for the software developers to use R520 architecture fully. This one will take time to complete. Anyway this will be achieved to succesfully use WGF 2.0 compatible hardware - and R520 architecture is closer to that than G70.
10-08-2005, 02:55 AM
slider99

Quote:

Originally Posted by Cooper

I agree with UnG on inefficient hardware use.
Lithan, actually I pulled this number out of my finger - I`m really confused with all those reviews. IMO non of todays alpha games have the ability to fully use R520 hardware.
But I think with the right software/drivers X1800 could be more than 20% efficient against GTX. It`s just for the software developers to use R520 architecture fully. This one will take time to complete. Anyway this will be achieved to succesfully use WGF 2.0 compatible hardware - and R520 architecture is closer to that than G70.

You mean like not implementing full SM3.0? Leaving out essentials such as "Vertex Texture Fetching"? Perhaps, Epic ought to recode their whole U3 engine to make it "more efficient" on ATI's hardware? :rolleyes:

It will be truly interesting to see what nVidia does next (81.xx drivers + 512MB 7800 + ?). I can agree that fear was really a fright.
10-08-2005, 04:54 AM
n00b 0f l337

Ha! Yes! Truth be revealed finnaly! I think this is were its paying off for intel. r520 may not be the best but there technology, like 90nm and maybe 65nm coming up and there other architechtures are amazing.

Also look at the graph, the 7800gtx is the only thing that works better under the full application over just kernel....
10-08-2005, 05:00 AM
One_Hertz

Quote:

Originally Posted by Lithan

One hertz, cooper, where are you getting these numbers from? I haven't seen any benches yet and really want to see some.

Yeah I just kinda looked at the reviews and made my own assumption according to the games that I play. Will be different for every person.
10-08-2005, 05:36 AM
Ugly n Grey

Quote:

Originally Posted by slider99

You mean like not implementing full SM3.0? Leaving out essentials such as "Vertex Texture Fetching"? Perhaps, Epic ought to recode their whole U3 engine to make it "more efficient" on ATI's hardware? :rolleyes:

It will be truly interesting to see what nVidia does next (81.xx drivers + 512MB 7800 + ?). I can agree that fear was really a fright.

The vertex texture fetching issue is a non starter, it's already been explained in other threads how ATI has built the work around for that. Until we see tests of it there no way to see if IQ is impacted. To me the issue is a nothing event, the ATI card will run all those those games just fine and I'm willing to bet they will run fast.

A company that released a game that didn't run on ATI cards or Nvidia cards would be a stupid company indeed.

Almost makes me wish I played games so I could join in the fanaticism over subtleties that make no real difference. This goes for the IQ complaints against Nvidia and the whining about the SM 3.0 ATI functions. In the end they get fixed and the games work. Then somone benches them and everyone is happy all over again and the bandwagon gets full again... sigh

In any case, the real question is how fast will the different architectures calculate these science packets. I certainly hope we can use GPU's from both companies to fold for team 36362 in the near future. Think about the numbers I could put up on a system with 4 875 Opterons and a pair of XL or XT's!!! :slobber: :slobber: :slobber:
10-08-2005, 07:33 AM
perkam

Quote:

Think about the numbers I could put up on a system with 4 875 Opterons and a pair of XL or XT's!!!

Yea, but Nvidia's GPU was never designed for this type of folding and I'm sure there are other types of calculations it can do faster.

Exhibit A: 7800GTX doesnt lag behind as much in every test

http://img192.imageshack.us/img192/4...0sgemv14lx.jpg

Exhibit B: This has more to do with ATI's Architecture and less to do with whether the X1800 is faster than the 7800 (even the x800xtpe ends up faster in these calculations)

http://img192.imageshack.us/img192/2...0slides9su.jpg

Exhibit C: In ATI's defense, the X1800 could show even more performance seeing as how it isnt optimized for the X1800's new "scatter" feature

http://img192.imageshack.us/img192/3...ptimize3py.jpg

I just wish this defeats the need for a ppu in the future if companies continue to keep making hardware with 3 to 4 times the processing power of today's cpus.

Perkam
10-08-2005, 08:17 AM
Shadowmage

Quote:

Originally Posted by slider99

You mean like not implementing full SM3.0? Leaving out essentials such as "Vertex Texture Fetching"? Perhaps, Epic ought to recode their whole U3 engine to make it "more efficient" on ATI's hardware? :rolleyes:

It will be truly interesting to see what nVidia does next (81.xx drivers + 512MB 7800 + ?). I can agree that fear was really a fright.

lol

Go read my post in that thread you're talking about.
10-08-2005, 08:41 AM
Ugly n Grey

Or imagine equipping your folding rig with a few of THESE Clearspeed devices...

http://www.clearspeed.com/news/pr.php?pr=28

But I am hung on using video cards, as most of the enthusiast community HAS video cards they do not use 24/7
10-08-2005, 08:42 AM
slider99

UnG > I generally agree, and it was not my intention to display severe symptoms of either fanatism or nvidiotism... However, one has to remember that Epic really is a TWIMTBP company, so I wouldn't be too sure. Moreover, I'm not sure what other threads you're referring to, because it's still very much unclear in those in the News section (smells big coverup from ATI, although it isn't an absolute requirement for SM3.0 - I give you that one). However, my reply was an answer to Cooper's intriguing suggestions in his posts...
10-08-2005, 09:02 AM
Ugly n Grey

Quote:

Originally Posted by Mikesta

Athlon 64's stack up badly against P4's in F@H. This is because F@H uses Intel compiler that does not allow SSE, SSE2 & 3 to be used with non-intel processors.

Just force SSE on at the command line. That's close enough for me. Darned if I will go out and buy Intel chips to fold with specifically, I'll use whatever is lying around. Be nice to have the video cards (even lesser cards.... 6800's, 9800's etc) be able to fold when not in use.

No worries Slider, it's just that in this particular case of folding with the chip I don't really care which card is the better gamer :) . The two companies use their pawns to hype up issues on forums like ours and get the guys fighting amongst themselves and it depresses me a little.

To no one in particular: If you are being funded by ATI or Nvidia directly or indirectly, you should step up and say so everytime you post...ya hosers. Get some ethics.

I'm going to get some time on the stanford DC forums in and learn a little more about the folding on video card way of the world. It seems the most promising..

GO TEAM 36362 :toast:
10-08-2005, 09:15 AM
nn_step

all this means to me, is I just need to put that card in my folding rigs but I can keep my Game box pure...
10-08-2005, 10:09 AM
Pinnacle

Cool stuff.

Folding Power!
10-08-2005, 11:26 AM
perkam

Shadow, pls stay on topic, this thread is about X1800XT's potential folding capabilities.

Perkam
10-08-2005, 12:39 PM
slider99

So, what makes for these excellent folding capabilities of the R520? Must be some extreme efficiency, considering less pipes and just 50% higher clocks. Possibly also the superior mem controller with 512-bit internal capabilities as well as the extra memory? When will GFX cards start help out with crunching? (Was never that into folding anyway.)

UnG > Yes, it's probably not too far-fetched to assume both nvidia and ati reps have a finger or two in all of this mess. However, they wouldn't probably know any more about the who is who than the next guy is, so I say, just let them fight it out while everyone else are enjoying their brand new XGI card. :D
And that other thread seems quite closed btw...

Fairly OT, but...
Shadowmage > Dil maintains that the feature's used for every texture in every frame in U3, and (assuming he hasn't been hallucinating too much lately) I wouldn't underestimate its impact on IQ and performance. But, as for now, we can only speculate, so I'm not sure this is a cause worth spilling too much blood for.

edit: well... seems like the black panther's been out hunting again. guess they don't want us discussing this issue, shadow. (*cough*coverup*cough*)
10-08-2005, 12:45 PM
rcofell

You mean X1800XT, right? ;)
10-08-2005, 01:06 PM
perkam

Quote:

Originally Posted by perkam

Yea, but Nvidia's GPU was never designed for this type of folding and I'm sure there are other types of calculations it can do faster.

Exhibit A: 7800GTX doesnt lag behind as much in every test

http://img192.imageshack.us/img192/4...0sgemv14lx.jpg

Exhibit B: This has more to do with ATI's Architecture and less to do with whether the X1800 is faster than the 7800 (even the x800xtpe ends up faster in these calculations)

http://img192.imageshack.us/img192/2...0slides9su.jpg

Exhibit C: In ATI's defense, the X1800 could show even more performance seeing as how it isnt optimized for the X1800's new "scatter" feature

http://img192.imageshack.us/img192/3...ptimize3py.jpg

I just wish this defeats the need for a ppu in the future if companies continue to keep making hardware with 3 to 4 times the processing power of today's cpus.

Perkam

Yes that was a typo rcofell, I meant X1800. But If you look at my post above rcofell, even the x800xt wins over the 7800GTX sometimes...so its a matter of architecture.

Perkam
10-08-2005, 02:54 PM
Ugly n Grey

A little birdy told me those numbers had been understated, the actual optimized results will significantly higher....

Fold baby fold..... daddy needs soem new PC's :)
10-09-2005, 02:05 AM
Mikesta

Quote:

Originally Posted by Ugly n Grey

Just force SSE on at the command line. That's close enough for me. Darned if I will go out and buy Intel chips to fold with specifically, I'll use whatever is lying around. Be nice to have the video cards (even lesser cards.... 6800's, 9800's etc) be able to fold when not in use.

You still can't get a64's to run with SSE2 on cores like QMD which are the most efficient cores to get points/time wise. Even on 364 pointers (The best PPW cores for a64's) a64's get outgunned by Intel CPU's. It sorta makes me sick. A64's would kick some major ass if they would use SSE2.

Anyway back to the topic at hand....ATI will have my money when I decide to upgrade my 6600GT because of the fact that it Folds so much better. The question I guess is...why is it so much better then Nvidia products?

Oh and on the clearspeed topic. I will be getting one for sure. What is great about it is that it's 50Gflops of processing power @ 25W. A big thing with hobby folders is the electricity bill. The Clearspeed math coprocessor has around 4 times +, the power of a good P4's which would consumer over 400W+ of power.

I just hope that the Clearspeed math co-processor will be supported or be able to be used with programs for other uses like encoding.

But back to the R520, around 25Gflops of power, unoptimised in Folding Gromac's is just great. I am really looking forward to the r580 hoping it will have a 512bit bus. 50+Gflops of folding power would be the equal of the clearspeed at around the same cost but have other uses (Crossfire!). Unfortunately they will probably consume 80-100watts of power.
10-09-2005, 04:53 AM
Ugly n Grey

Yeah, I was thinking the same thing. I don't play games but I'd buy one to fold on. As far as the SSE2 fiasco and the QMD packets, I have some of both camps products (A64/P4), so like I said, that good enough for me (at least for now). It won't be long before there's a 64 bit edition as well which should help both camps.

Well the next gen cards alone should match the Clearspeed dual chip cards if they can get the code optimized. I think the clearspeed is an easier product to code for, but I won't buy one. It's a limited use product to me. At least on a video card I have a display function :D which I DO need.
10-09-2005, 11:42 PM
STEvil

Heres two threads on the hacked intel compiler(s).

Cant find the 3rd one, think it was started by Lithan but I could be wrong.. let us know if you figure out how to use this with F@H and how it impacts perf. ;)

http://www.xtremesystems.org/forums/...hlight=compile
http://www.xtremesystems.org/forums/...hlight=compile
10-10-2005, 01:35 AM
Mikesta

hmmmmm im not too likely to try it.

Like Ugly n Grey , I am waiting for the 64bit patch. Apparantely it will play nice with A64's.

Anyway I hope Stanford gets off it ass and releases GPU client!

Even more so I hope that ATI gets off it's arse and gets its r520 or even r580 out as soon as they can!!!
10-10-2005, 07:37 AM
P_1

If some1 wants to try patching folding@home for amd 64 here is the iccout program to do it http://www.xtremesystems.org/forums/...chmentid=29223
originally posted by Walrusbonzo

If I had an athlon 64 here I would try it

Also what you need to do is recompile the program using the -xN flag first, then use this program to patch it, otherwise it probably wont work.
10-10-2005, 12:07 PM
saaya

Quote:

Originally Posted by perkam

Yea, but Nvidia's GPU was never designed for this type of folding

but atis vpu was?
10-10-2005, 12:40 PM
P_1

Quote:

Originally Posted by saaya

but atis vpu was?

Well one of the architects talked about using their vpus as physics proccessors as well, so i guess they had it in mind.
10-10-2005, 02:30 PM
angra

Quote:

Originally Posted by Ugly n Grey

Distributed computing programs are nothing new and neither are discussions about using the VPU of your video card to fold. A recently leaked document from the Stanford University Graphics Lab has the X1800XT leading it's rival the GTX by over 7x in certain types of calculations with performance equivelant to a 10.5 GHz P4 CPU.

I do a fair amount of work in similar areas to this group at Stanford. I can tell you with near certainty that I can find a computation kernel that has a 7800GTX beating a Radeon X850 by 7x. The programming model, performance bottlenecks, and data-access and flow restrictions are so bizarre, and variable from GPU to GPU that disparities like this are really not particularly unusual.

As far as folding goes, I think that GPUs are pretty well suited to folding. In fact I know a guy building a custom GPU-based cluster to do private folding. Kind of neat. Anyway. I dont think the stanford folding guys open source the folding algorithms, so someone with access will have to rewrite the computational kernels to take advantage of GPUs. Fortunately, Stanford also has many of the pioneers of GPGPU.
10-11-2005, 02:53 AM
Le_Petit_Lapin

What if you get 2 and crossfire them, and then run a CPU based folding client as well?

Would that be pure folding ownage?
10-11-2005, 03:04 AM
Mikesta

Quote:

Originally Posted by Le_Petit_Lapin

What if you get 2 and crossfire them, and then run a CPU based folding client as well?

Would that be pure folding ownage?

ABSOLUTELY!

But watch the power bill....

Then again a system with 2x 512mb X1800 will do the work of many systems (folding boxes) and will be much more power efficient points wise.
10-11-2005, 04:22 AM
Ugly n Grey

agreed on the power use, multi core, multi card systems are more efficient that way. I think this is going to be fun, folding on cards especially at a place like xtremesystems where everyone and their uncle owns something fast!
10-11-2005, 07:30 AM
FireDragon

Now think of the XBOX 360 and folding with an even faster core than R520 and what is it 3 or 4 3.2 ghz g5's?

Dragon
10-11-2005, 12:34 PM
freecableguy

Since the X1800XT isn't available yet that wouldn't make it faster at anything....
10-11-2005, 01:51 PM
twiggy

Quote:

Originally Posted by freecableguy

Since the X1800XT isn't available yet that wouldn't make it faster at anything....

:rolleyes: :sick:
10-11-2005, 02:07 PM
nn_step

now I wonder if you could set up 4 x1800xts in a Quad Sli board with a Dual core proc for insane folding...
10-11-2005, 03:27 PM
Ugly n Grey

If you were JUST buying cards to fold with, The clearspeed option looks easier to manage technologically really. Mind you, I'm sure you could find another use for a quad card system occasionaly :slobber:

Buy a video card... fold for XS :)
10-11-2005, 05:12 PM
nn_step

Actually I am folding for 2 other teams, in order for me to add XS folding to my army of boxes, I would have to buy another couple rigs..Christmas would be as soon as I could add a couple more...
10-11-2005, 05:27 PM
Ugly n Grey

We just ask folks to add what they can, we need all the help we can garner. A couple guys in the team have temporarily given us a HUGE boost, Xciter gave us the use of 330 + CPU's for a while and Angra has been pushing points to us with an extremely sweet Linux cluster. Hell of a pile of production out of only a few users
10-12-2005, 01:47 PM
pythagoras

Quote:

Originally Posted by angra

I do a fair amount of work in similar areas to this group at Stanford. I can tell you with near certainty that I can find a computation kernel that has a 7800GTX beating a Radeon X850 by 7x. The programming model, performance bottlenecks, and data-access and flow restrictions are so bizarre, and variable from GPU to GPU that disparities like this are really not particularly unusual.

As far as folding goes, I think that GPUs are pretty well suited to folding. In fact I know a guy building a custom GPU-based cluster to do private folding. Kind of neat. Anyway. I dont think the stanford folding guys open source the folding algorithms, so someone with access will have to rewrite the computational kernels to take advantage of GPUs. Fortunately, Stanford also has many of the pioneers of GPGPU.

If you would be interested in collaborating on this I could talk to vijay, pm me to let me know.

Regards

John.
10-12-2005, 03:15 PM
angra

Quote:

Originally Posted by pythagoras

If you would be interested in collaborating on this I could talk to vijay, pm me to let me know.

Regards

John.

actually after I said that, I looked around some more and see that the stanford guys are already working on this. They are apparently working from the BrookGPU direction, which makes sense for them. Brook is a pretty good langauge and maps pretty darned well to GPUs. So I think the best answer is to sit back and enjoy the Stanford genius-show.
10-12-2005, 09:50 PM
Lithan

I've moved some of the discussion regarding d2ol into a seperate thread which will temporarily reside in an invisible forum until it's decided where to put it. If interested in d20l please visit the forums about it at... http://www.xtremesystems.org/forums/...play.php?f=159

Other than that, please try and keep discussion in this thread directly related to this news item.

Thanks.

All times are GMT -8. The time now is 01:59 AM.

XtremeSystems