As I already mentioned in some other posts, I am working on an implementation of the MW application on ATI graphics hardware (RV670 and up, that means HD38x0, HD4670?? and HD48x0). It is not finished yet (still some scheduling problems as BOINC only supports CUDA atm), but it calculates some results that are in line with the requirements (at least when running offline, have to ask Travis for this ones).

But I guess you want to know how fast it is! So without further fuss here is a screenshot for you (done on a PhenomII 940BE with a Radeon HD4870):



That's something I would call fast!

Holy mother of moo moo's... 8seconds per WU!!

Great to see some people can develop such an application on a GPU (not the easiest of things!) Just a shame that the BOINC dev's have decided to cast ATI to one side and adopt only CUDA is all i can say.

Well done, keep it up :)

I agree, excellent work!!

I know nvidia cards aren't as fast as ATI for this kind of work but when you have completed the ATI app would you consider a CUDA app?

I've just had a go at knocking out my first GPU WU which recently completed giving me nearly 4,000 credits.

I was very impressed, but much more so with the fact that all the while my laptop was churning out 2 MW WUs every 8 minutes or so, at the same time.

I knew VM would eventually let me squeeze out more.

But this is all nothing compared to how I now have an optimized signature with 6 images and 5 links. Who can beat that ? :P

---

I've just had a go at knocking out my first GPU WU which recently completed giving me nearly 4,000 credits.

I was very impressed, but much more so with the fact that all the while my laptop was churning out 2 MW WUs every 8 minutes or so, at the same time.

I knew VM would eventually let me squeeze out more.

But this is all nothing compared to how I now have an optimized signature with 6 images and 5 links. Who can beat that ? :P

Travis, please transfer those 4,000 credits to my account. That beats all.

---

me@rescam.org

Now there is the making of a conundrum ... which card to get ... ATI so you can do Milky Way or Nvida so you can do SaH and / or GPU Grid?

Or different cards for different systems ... :)

Just a side note, it is possible that GPU Grid got the CPU load issue under control, at least at first blush ... I have one task in flight right now with 1% tops CPU load ... over 10% done in two hours and just over a minute of CPU time ...

---

w00t, sign me up! Just been waiting to get my 4870 1GB rolling. I was waiting on AI because they're working on an official app last I heard, but Milkyway will work too! :-D

---

That's really crazy. Now I wish I had an ATI card...
But:
Does this break the credit limit barrier of 108? Wonder if.
Otherwise it's only another method to crunch more WUs without getting more credits from it.

---

Member of BOINC@Heidelberg and ATA!

My BOINCstats

Otherwise it's only another method to crunch more WUs without getting more credits from it.

You mean, like work without pay?

---

Otherwise it's only another method to crunch more WUs without getting more credits from it.

You mean, like work without pay?

Not without pay, but close to it!
With the current limits (especially the 300WU limit per core), you get only about 320 credits/day on that system with the GPU app *lol*

If it is really going to work, Travis has to do something on this.

Not without pay, but close to it!

Gulp! I can't image anyone providing computer resource to a BOINC project like this and getting almost negligible credit results in return.

---

Not without pay, but close to it!

Gulp! I can't image anyone providing computer resource to a BOINC project like this and getting almost negligible credit results in return.

Just think of it as a small fan-heater that supplies 8k+ MW units/day when its not needed as a graphics card.

If someone has thought, the screenshot may be a fake, there is a video of the running app.

I've implemented a solution to the scheduling problem (at least for Windows) that seems to be working on that specific machine. It is able to calculate several WUs at once, but that doesn't raise the throughput of the app, so calculating 4 WUs at the same time on a quadcore isn't faster than one after another. It uses up one core independent on the number of concurrent WUs.
In principle it works like a small scheduler that switches between the WUs 20 times a second (HD4870, 6 times a second on a HD3850). The video shows that GPU-Z is somehow incompatible with this approach as one experiences severe lags or short freezes of the system after it is started.

Thanks Emploi for that video!

If someone has thought, the screenshot may be a fake, there is a video of the running app.

I've implemented a solution to the scheduling problem (at least for Windows) that seems to be working on that specific machine. It is able to calculate several WUs at once, but that doesn't raise the throughput of the app, so calculating 4 WUs at the same time on a quadcore isn't faster than one after another. It uses up one core independent on the number of concurrent WUs.
In principle it works like a small scheduler that switches between the WUs 20 times a second (HD4870, 6 times a second on a HD3850). The video shows that GPU-Z is somehow incompatible with this approach as one experiences severe lags or short freezes of the system after it is started.

Thanks Emploi for that video!

Jeeeeeezzzz... And i thought you mean't 1 result every 10 or so seconds! But 4!! Wohay, why the hell did i get Nvidia! :P

Great app!

Jeeeeeezzzz... And i thought you mean't 1 result every 10 or so seconds! But 4!! Wohay, why the hell did i get Nvidia! :P

Great app!

It's 4 WUs every 35 to 40 seconds or so. It is effectively every 9 seconds a WU.

If you lower the resource share (calculate some other projects in parallel), the app will calculate less MW-WUs in parallel (you can run other projects in parallel), but the absolute troughput will stay the same (1 WU per 9 seconds).

w00t, sign me up! Just been waiting to get my 4870 1GB rolling. I was waiting on AI because they're working on an official app last I heard, but Milkyway will work too! :-D

I have an ATI Radeon X1900 in my Mac. Two questions.

1) "...that means HD38x0, HD4670?? and HD48x0" What the heck does that mean? Which cards/chips are we talking about?

2) Will this app work on Macs?

---

Two questions.

1) "...that means HD38x0, HD4670?? and HD48x0" What the heck does that mean? Which cards/chips are we talking about?

2) Will this app work on Macs?

1. Consider the x as a wildcard. It runs on RV670 and RV770 class GPUs (and anything newer supporting double precision, so it's forward compatible). That are HD 3850, HD3870, HD4850 and HD 4870 (most probably HD4830 too). As for the RV730 (HD4670 and HD4650) there is some contradicting information. It is tested not to run with the current driver, but this may or may not change with a newer driver revision.
2. Macs are not supported (you can't even download a driver for the newer cards from AMD, so I suppose they wouldn't run). But there is potential support for Linux (not implemented yet).

[quote]

I have an ATI Radeon X1900 in my Mac. Two questions.

1) "...that means HD38x0, HD4670?? and HD48x0" What the heck does that mean? Which cards/chips are we talking about?

2) Will this app work on Macs?

your X1900 will not work. you need at least a card with the RV670 chip (3850, 3870) or with RV770 chip (4830, 4850, 4870).
In principle all ATIs starting from the RV670 should work, but sometimes the double precision is locked by either the bios or driver (4670 refused to work)

8 tasks on 1 gpu ;)

http://de.youtube.com/watch?v=nnsW-zB95Is (new video)

Is there any way to get the application now, or is it just open to Planet 3DNow members?

---