Stream

[Brucifer]

Has anyone been messing with the stream beta? If so, what are you using it on?

[Ototero]

Not me, I'm innocent. :)

[Brucifer]

I got my stream going................. :icon_mrgreen: Pun intended, I know, I know, very sic humor. :icon_mrgreen:

Anyway, on with the story. I noticed yesterday that distributed.net had released a 32-bit windows stream client for beta8. So I broke down this morning and raided my coin jar where I put all my pocket change every evening. And there was enough in there to go out and buy an ATI (AMD) Stream card. Unfortunately the store was out of the 48xx that I wanted, so I ended up having to settle for a Radeon HD 4650. I figured that it would suffice to get my interest in the effort whetted a teensy weensy bit. All my other stuff is cuda. So I was curious as to how the Stream stuff would compare on a few levels. As for ease of installation. I'm running XP Professional on this system. So the first thing I did was download the beta8 stream client. Then I went and did a google on ATI Stream, and lo and behold there was a "suite" of stuff from ATI so I downloaded that. Then I replaced the cuda card (and put that in an XP-64 system) and then installed the 4650. Fired the system off, and clicked on the "suite" and let the installation go forward. Afterwards a note popped up and said that I needed .NET 2.x installed. So I went to MS and downloaded the latest .NET and installed that, along with a security update or two.... :icon_rolleyes:

Then I installed the stream beta client. Fired that off in config and put in an email addy, and aimed it at my internal perproxy server. Then I rant the -bench util to see how the timings went. On the plus side, the client picked the correct default core for speed, however there were only 2 to pick from. LOL Then I fired the client off. It takes 24 secs per work unit, and I'm only running single work units at the moment. For comparison's sake and not to get into a cuda vs stream war here, a cuda 2.2 client on a 9800 GTX+ PNY mfr cardtakes 21 secs per unit. They are fairly close together. The cuda one does better on no screen lag when using the system while crunching. The ATI lags noticably, but not to the point where the system can't be used for stuff like forum typing, etc. The ATI was $129 and the cuda one was 147 out the door tax included.

So at this point it's pretty much a wash between the two cards. The ATI is a few bux cheaper and is fine for a cruncher system. Where the ATI card does better is that it doesn't need power plugged into the card. It is totally quiet and isn't giving off much heat at all. The cuda card is much larger and requires the additional power plug. I do have to say that I am impressed with the little ATI card though as the installation went flawless. Of course it isn't going to be flawless on linux............. but then the cuda cards aren't either cause with either you are playing the kernel compile game with SDK's and all. Performance wise they run the same really. So what do you gain with linux? Savings on the cost of the O/S is about it. With the beta game that distributed.net is going through working on the clients, the windows method is much much simpler for doing the GPU thing as far as RC5.

The larger ATI cards (which I would have bought a larger one but the store was out and I needed instant gratification) do much much better on speed. So after going through the installation on this one, I won't have any hesitation on spending the bucks for a higher end ATI when my pocketbook allows again.

But my main thing in posting this is that it is an AMD GPU card, and since this is the AMD users forum, it is nice to finally see AMD getting into the real world in the GPU thing for doing work and not just the gaming thing.
So if I didn't have to pay state sales tax here, the card was $119 at the local Best Buy, and cheaper at NewEgg. So if you have a hankering to get into playing with ATI stream on RC5, this is a relatively inexpensive way to go. This card does better than a GeForce 9800 GT by 4 seconds per work unit. So anyway, as far as distributed.net, ATI is in the game with cuda.

[Nflight]

Woo Hoo Thanks for posting Brucifier, I can finally believe again in AMD ! WoW I feel better for a while there it was toting the barge trying to support the AMD line of GPU's and Processors, they work sure but, just not the performance we all wanted to see. Now they seem to have emerged once again victorious. :blob3: :hello:

I feel Better, thanks for commenting...!!!!

[Brucifer]

I lifted the below from the Free-DC rc5-72 forum thread, and the post was by Digital Parasite:
paste-----------------------------
[Aug 16 00:11:25 UTC] RC5-72: using core #0 (IL 4-pipe c).
[Aug 16 00:11:35 UTC] RC5-72: Benchmark for core #0 (IL 4-pipe c)
0.00:00:07.42 [589,471,423 keys/sec]

With a Core2 Q9550 and an ATI 4870 (GPU 750 MHz, Memory 950 MHz) running Vista 64bit.
paste--------------------------------

So the 4870 was what I was wanting but the store didn't have any. What had my heart started there was the benchmark of 589,471,423 keys/sec which is flat rocking along... So it puts out some heat, well they all put out the heat when you get into the stuff that is putting out 400,000,000 keys/sec and up. So I want one of those puppies to play with. That puts AMD right up in the thick of it as far as rc5072 goes!!!

[Brucifer]

A bit of an update. Actually the stream is better than in the thick of it. The HD 4870 is a real ball buster on rc5 and for the price there is nothing else on the market that will touch it in performance. So AMD finally has a real winner on it's hands. So it stands to reason that if it is such a hot performer on rc5 it's going to do the same on other projects too. So it's good to see AMD getting back into some serious competition again. :)

Now I just gotta find a place that isn't out of stock................... :icon_twisted:

[NeoGen]

And if the boinc developers pick it up and start working on Stream, there will be no stopping. :)

If only ATI's Stream was being as well marketed and supported as CUDA is...

[Brucifer]

Give it a little time. Cuda has been on the street a while. :)

[Brucifer]

For anyone that's interested, we've had a bit of a discussion going on the distributed.net threads of the FDC forums concerning benchmarks on some of the Nvida GeForce and AMD Stream GPU's.

[Brucifer]

I picked up a Radeon HD 4850 last night, and the benchmarks are;

[Aug 21 15:35:18 UTC] RC5-72: using core #0 (IL 4-pipe c).
[Aug 21 15:35:29 UTC] RC5-72: Benchmark for core #0 (IL 4-pipe c)
0.00:00:08.84 [489,096,269 keys/sec]
[Aug 21 15:35:29 UTC] RC5-72: using core #1 (IL 4-pipe c alt).
[Aug 21 15:35:42 UTC] RC5-72: Benchmark for core #1 (IL 4-pipe c alt)
0.00:00:10.59 [411,941,905 keys/sec]
[Aug 21 15:35:42 UTC] RC5-72 benchmark summary :
Default core : #0 (IL 4-pipe c)
Fastest core : #0 (IL 4-pipe c)

So it moves along at better than twice what I'm getting out of a cuda BFG GTX 260. Uses one PCIE 6-pin power cable, not two like the GTX 260. Feels like its putting out about the same (maybe less) heat than the 260. It crunches one single work unit each 8 seconds, so it's like rocking along like the cuda 260 was before the compiler problem hit. And it's not the fastest card of the stream line, but I'm pretty happy with the way it runs. :icon_mrgreen:

So there you go Nflight, your AMD world is now in the thick of it in the GPU world as far as crunching on RC5-72 anyway!!!

[Nflight]

Bok made this statement which will rock the world of AMD. 8-18-09


They've just announced that they should have the ATI integration into the BOINC 6.10.x stream within a month, it's basically the major thing they are working on now that the gridrepublic/facebook 6.8.x version is out the door..Woo Hoo :blob3:

[Brucifer]

yup, these cards rock!!!

[liuqyn]

boinc 6.10.0 is out now (beta), but I can't get it to run on my AMD systems, runs fine on my only intel(P4 2Ghz) though.

[NeoGen]

I'm running it here (the beta) on AMD Phenom II X3 720, just upgraded from 6.6.36 :)



23/08/2009 04:07:29 Starting BOINC client version 6.10.0 for windows_x86_64
23/08/2009 04:07:29 log flags: task, file_xfer, sched_ops
23/08/2009 04:07:29 Libraries: libcurl/7.19.4 OpenSSL/0.9.8k zlib/1.2.3
23/08/2009 04:07:29 Data directory: C:\ProgramData\BOINC
23/08/2009 04:07:29 Running under account NeoGen
23/08/2009 04:07:29 Processor: 3 AuthenticAMD AMD Phenom(tm) II X3 720 Processor [AMD64 Family 16 Model 4 Stepping 2]
23/08/2009 04:07:29 Processor: 512.00 KB cache
23/08/2009 04:07:29 Processor features: fpu tsc pae nx sse sse2 pni
23/08/2009 04:07:29 OS: Microsoft Windows Vista: Ultimate x64 Edition, Service Pack 2, (06.00.6002.00)
23/08/2009 04:07:29 Memory: 3.99 GB physical, 8.21 GB virtual
23/08/2009 04:07:29 Disk: 148.08 GB total, 106.61 GB free
23/08/2009 04:07:29 Local time is UTC +1 hours
23/08/2009 04:07:29 NVIDIA GPU 0: GeForce 9600 GT (driver version 19038, CUDA version 2030, compute capability 1.1, 512MB, est. 31GFLOPS)
23/08/2009 04:07:29 No CAL Runtime Libraries installed.
23/08/2009 04:07:29 Not using a proxy
23/08/2009 04:07:30 Version change (6.6.36 -> 6.10.0)



I highlighted in red the part that seems to be what identifies ATI GPU's :)

[liuqyn]

but does it run on cpu? when I installed it, it keeps setting number CPU is 0. and no matter what I change in preferences it stays that way.

[BobCat13]

but does it run on cpu? when I installed it, it keeps setting number CPU is 0. and no matter what I change in preferences it stays that way.
In your BOINC Data directory, open the cc_config.xml file. Find the entry like this:

<ncpus>0</ncpus>

Change the 0 to -1, save the file and either stop/start the boinc client or in BOINC Manager choose Advanced/Read config file.

Earlier versions of boinc ran even though that setting is 0, but 6.10.0 will set usable CPUs to 0. -1 tells the client to use all availabe CPUs.

Also, be aware that on newer versions of boinc, the checkpoint (Write to disk at most every) setting is multiplied by the number of CPUs on the machine. If you had set it to 5 minutes (300 seconds) on older versions, it will now save checkpoints every 10 minutes on a dual core, 20 minutes on a quad, 40 minutes on an octo. This can cause a fair amount of lost work if a reboot or some other restart of boinc is required.

[liuqyn]

ahhhh, that did it. many thanks.

ps. sorry for hijacking the thread.:icon_wink:

[Nflight]

Speaking of hijacking a thread, just which is the AMD ATI GPU's that are referred to as Stream Capable. I would like to know?

EDIT: I would also like to know if you can run separate BOINC Projects while the cpu's are crunching the GPU's are crunching something else?

[NeoGen]

EDIT: I would also like to know if you can run separate BOINC Projects while the cpu's are crunching the GPU's are crunching something else?
That is possible, but you have to go to the website and set one project on the preferences to use only gpu and the other only cpu. (or leave as is if no gpu apps)

[liuqyn]

most of the HDxxxx cards are stream capable, however, milkyway@home which was the first boinc project to use ATIs requires the double precision capable cards which is the 38xx and 48xx series models.

and yes you can(and actually have too) run different projects as the ATI optimized apps run CPU tasks via the anonymous platform in boinc which make it impossible to also run regular CPU tasks at the same time for the same project.

check out Milkyway (http://milkyway.cs.rpi.edu/milkyway/) and Collatz Conjecture (http://boinc.thesonntags.com/collatz/index.php) message boards for more info.

[Jason1478963]

Very interesting to see some projects supporting these video cards. I would like to see some boinc projects doing this as well. It seemed pretty straight forward on the install. I'm just not to sure on the stats and user account. It looks like I have a while to wait to get added to the stats so I can edit my profile. I watched my GPU hit 70C when it was running over 97percent activity with the fan at 100 percent. It looks like I would need a liquid setup to keep this 4850 cool at 90+ GPU activity

Thanks for all the great info on this project Brucifer. It looks like your gonna be moving up the stats pretty fast hitting number 18 on the daily stats today.

[Brucifer]

I just let them run. The 4850 isn't running any warmer than the GTX 260, and it puts out double the work on rc5, and much less current draw too. It's a good card, and I will buy more of them I think and get rid of the other stuff I've used for rc5 and save a bit of electricity too. :-)

@liuqyn -- what are you running on rc5??

[liuqyn]

ran a little on my laptop last week(IGP HD3200 roughly 5-6 Mkeys/sec), then one batch of cuda on my GTX285 triplets (combined roughly 700-800 Mkeys/sec). currently not running.

[sigma7]

Hi

Found this board by googling, one of the few places I can find where people are discussing RC5-72 and the graphic accelerators to power them.

I'm having some amazing results from the new 2.9108-517 client (released mid-March 2010) versus the 2.9108-516 client.

I am running a stock 5850 and have seen my keyrates increase from ~966Mkey/s to ~1,420Mkey/s. The release notes for the client indicate it now has a faster core for HD5xxx cards so I'm guessing all sorts of benefits could be had even on the cheaper units. Given I'm getting a 50% boost, that's pretty impressive.

For reference, the GTX260+ (factory overclocked) in my old PC still only manages a (relatively) poor 245Mkey/s. Although after years of crunching on the CPU and thinking 15Mkey/s was awesome, it does seem weird to criticise the GTX260+ for its 'weak' performance.

[Brucifer]

LOL, yes I hear you on the GTX-260. I've got one sitting on a desk top because they suck so much juice and put out a ton of heat. I'm running a 5770 in place of it. It isn't super fast either but it sure doesn't put out the heat and is twice as fast as the 260.

Now if they would finish their testing on the database and get it back read write, so I could get my primary rc5 account back under the team, things would be great. :)

[Brucifer]

The problem with the newer cards is that the prices of them just shoots way up beyond reason, and then a year or so later it is like a third of the initial cost. I have one lowly XFX 4650. It doesn't put out like the big ones, but it cranks out about 2,880 units/points a day running on a 450w power supply so it doesn't suck down much juice. That's like 1,051,200 units a year out of that one little card. Still just knocks the socks off any quad out there, and much much cheaper to purchase too. :-) And while it doesn't put out like my 4850's, it is still a pretty marvelous item when I look back several years to when I was running a herd of processors and getting much less that the 2,880 units a day. We get pretty spoiled on performance now days and forget what the performance leaps are in a short time span.

My 4850 runs 9,600 units a day or 3,504,000 or so units a year, over three times what the xfx 4650 will do. So those gpu's are definitely worth their weight in gold when viewed from a cost versus output perspective. Much cheaper than buying new cpu systems. So they are the way to go for crunching rc5.

My GeForce cuda systems on the other hand are much slower than the ATI systems and suck down much more electricity too. However when compared with what the cpu systems put out, the present cuda systems still are impressive items. It will be interesting to see what the next generation cuda stuff does, and what Nvidia is going to charge for them too.

[Brucifer]

I have taken my rc5 crunchers and put them on collatz for the time being. When rc5 gets back to allowing the user database to be edited then I may bring them back here, unless of course the dark side gains control of my dc addiction....... :icon_rolleyes:

[drezha]

Sigma,

I've found the completely opposite. I was using the older client on my 3850 (or whatever I have - I forget!) and it would fly along but with the new release, it's broekn and I get nothing. Seems to download everything fine but no progress. Maybe I didn't leave it long enough but I with the new client, I cant use it at all.

Major shame.
Means all I get is about 50 units a day whilst at work (one core of a T9400 @ 2.53GHz as I don't want to overheat my laptop).

[Brucifer]

Are you still crunching with the older release somehow???? Cause that will not register any completed units. You must use the new release. Somehow you got them mixed up?????? The new release from the distributed.net download page works for both stream and cuda....

[drezha]

I download the new release (v2.9108.517) and I get nothing. I select a manual benchmark and it goes through that and chooses the IL-4 Pipe alt C.

EDIT:
Anyhow, whilst typing this, I set up another instance and it appears it takes about 3-4 minutes to get itself going. And now I have it's showing about 19,000,000 Key/s (which seems a bit pants considering my CPU at work kicks out 9,000,000 Key/s per core IIRC)

http://www.drezha.me.uk/gallery/albums/userpics/10001/2010-03-31_154105.png

As you can see, it troughs and peaks a lot :(