That's probably because I don't find their science very compelling ;)

Yes, well, you included SaH and they are not doing science at all ... they are doing exploration ...

The only point I was trying to raise is that like a particular project or not or find their purpose interesting or not, when we write summaries of the world of BOINC we should be as comprehensive as possible because someone else may find the project's focus compelling even when we think it is a waste of time.

Sorry if I wasn't clear: I'm grateful you corrected the list (or at least made it more complete). I didn't leave it out intentionally, I really forgot it. And my last statement was supposed to be taken humorously. I admit, not exactly my best post.. :p

MrS

---

Scanning for our furry friends since Jan 2002

This is a Fermi thread on GPUGRID. Lets keep it that way, please.

To keep this thread about Fermi cards ... I don't have one ... :)

Sorry if I wasn't clear: I'm grateful you corrected the list (or at least made it more complete). I didn't leave it out intentionally, I really forgot it. And my last statement was supposed to be taken humorously. I admit, not exactly my best post.. :p

Sorry, I missed the humor ... :(

My fault entirely ...

And I still don't have a Fermi card ...

While GPUGrid is clear on the fact that their app is not quite ready to run on a GTX480 I thought I would exercise my new card to get a ballpark feel for it. I will be running a couple other projects but I'm not going to do any *optimized app* installations as I am a simple kind of cruncher.

Test system is an i7-920 w/HT on @4.2 GHz running Win7 64, BOINC 6.10.43.
Asus P6T, EVGA GTX480, Corsair 2x3 1600 C7.

Milkyway does not work without an opti app.

SETI works good with shaders up to 1732 (1803 is a little shakey and 1823 grey screens).
It is hilarious to watch WUs crank through in 15-17 SECONDS instead of hours, the uploads and downloads can hardly keep up :big-grin:
SETI has some WUs that are calculating signals from low angles and they run slower on a GPU than on a CPU.
It's a known issue. There is an opti app that will switch to your cpu when these WUs are detected.

Colltaz: stock is good, 1532 is good, 1632 is good, 1732 good.
Collatz @ 1732 shaders is pushing the GPU usage at 85% which is about the same as GPUGrid does on my GTX285 and GTX295 cards with the same OS/CPU/RAM configuration.

With an ambient temperature of ~ 23-24c I manually set the fan at 70% and running temperature is ... 73c. Not quite the furnace we had been told to expect. This is in an enclosed full tower case that has excellent air flow (front, side, top, and back case fans). Everything looks pretty good to me so far, temperature and fan noise are well within my personal tolerance levels.
I have not put an extra 120 mm fan blowing on the backside (like I do on the GTX295) because it just doesn't need it.

When GPUGrid is ready, which I have every confidence will be soon, I will be switching back here where I believe I belong.

---

Thanks - Steve

SETI works good with shaders up to 1732 (1803 is a little shakey and 1823 grey screens).
It is hilarious to watch WUs crank through in 15-17 SECONDS instead of hours, the uploads and downloads can hardly keep up :big-grin:
SETI has some WUs that are calculating signals from low angles and they run slower on a GPU than on a CPU.
It's a known issue. There is an opti app that will switch to your cpu when these WUs are detected.

Are you sure (please check) that those 15-17 second SETI tasks are producing validated results?

There is a common failure mode at SETI where older CUDA cards suffer - I believe - memory corruption following an error, and return all tasks with:

SETI@Home Informational message -9 result_overflow
NOTE: The number of results detected exceeds the storage space allocated.

This is a valid outcome for some tasks (roughly 5%), but if you get it on every task, it's an error, and damaging to the project because of all those downloads.

But if you can link me to the host, and demonstrate that the tasks are validating correctly and getting full credit, I've still got time to nip into town and pick up a Fermi this afternoon... ;-)

Just checked in on the results I sent in earlier this morning.
Only one has been processed and it is in fact the Invalid situation you posted :-(
Fortunately I had already stopped my SETI testing so am no longer pushing the servers to feed me. I'll post back if any of the others validate properly.

---

Thanks - Steve

Just as well they were out of stock, then! "More expected on 28/04/10" - or 13 May, for some brands.

Edit - if you do want to test out the performance of the card on SETI tasks, the SETI Beta project has a Fermi-compatible application under test. Work supply is a bit erratic at the moment, though, because they're testing other things as well.

Some good news about Fermi here. They already shipped 400k units, which puts some weight behind their claim that the yield problems are solved by now. We should also see a 512 shader card at some point.
I'm still not interested in buying one, but an ultra high priced, unavailable Fermi wouldn't have done us any good.

MrS

---

Scanning for our furry friends since Jan 2002

My local Frys had about 6 of them on the shelves as a mix of 470 and 480s ...

Sadly, no 5870 or I might have been tempted to buy another ...

I would not read too much into xbitlabs interpretation of the financial state of NVidia.
Firstly, the cards only started to ship towards the end of the first quarter, so they could not have made a huge impact on first quarter finances!
Secondly, companies tend to offset losses from one quarter against others (even over a year), for TAX/Financial Stability reasons.

I was surprised to hear that a 448 cored Fermi popped up in the Tesla C2050, and that report sort of suggested they had sold a lot of these. Well you can be sure they made no money there; they are not due for release until Q3!
C2050

Similar here regarding Frys. (25yr anniversary sales was last weekend) I was shopping for a new computer for my kid. Basically, he wanted to run Starcraft 2 and I wanted a strong GPU to do cpugrid. I first tried to do a name brand computer (already built) but found nothing with video cards that were listed on cpugrid.

So, went 'clone'. I knew that I was going to make some heat, so got a Cooler master haf case. The only thing on the shelves that were listed on cpugrid was the gtx480. It was listed as 'recommended' on forum page. (I didn't realize till now that there are issues with it on gpugrid)

Similar here regarding Frys. (25yr anniversary sales was last weekend) I was shopping for a new computer for my kid. Basically, he wanted to run Starcraft 2 and I wanted a strong GPU to do cpugrid. I first tried to do a name brand computer (already built) but found nothing with video cards that were listed on cpugrid.

So, went 'clone'. I knew that I was going to make some heat, so got a Cooler master haf case. The only thing on the shelves that were listed on cpugrid was the gtx480. It was listed as 'recommended' on forum page. (I didn't realize till now that there are issues with it on gpugrid)

The are no issues with a GTX480. It is the fastest card you can buy for GPUGRID with a factor almost two faster than the other ones.

gdf

I've been watching the Fermis a bit trying to decide if it's worth buying one. So far most of the 470s around are running faster on comparable WUs than the 480s. Don't know why. I did find one 480 that was faster than any of the others and it's running at just a bit less credits/hour than a fast GTX 295 (which of course has 2 cores) and 1.63 times the speed of my GTX 260. So they are the fastest single core card, but not sure if the price is justified yet.

I've been watching the Fermis a bit trying to decide if it's worth buying one. So far most of the 470s around are running faster on comparable WUs than the 480s. Don't know why. I did find one 480 that was faster than any of the others and it's running at just a bit less credits/hour than a fast GTX 295 (which of course has 2 cores) and 1.63 times the speed of my GTX 260. So they are the fastest single core card, but not sure if the price is justified yet.

The price is not justified, it is clearly too expensive. But with cuda3.1 the performance is there. We cannot release the new version until 3.1 is out.
The GTX480 are way faster than the GTX470 (at least in Linux).


gdf

GDF is looking at his GTX480 running on Linux and using CUDA 3.1. Natively a GTX480 is 21% faster than a GTX470.
Perhaps you are looking at an exceptionally reliable and overclocked GTX295 and compairing it to one of several GTX480's on Vista or Win7 and suffering around 60% speed loss. This is driver related.

Most GTX295's only RAC between 40K and 70K per day.

My GTX470 is clocked at the same speed as a GTX480, and my card is on Win XP x86. It does have one thread left aside for GPU consumption, with swan_sync=0 in use.

It could (going by tasks) get 70K per day, and is moving steadily towards that (no failures on that system to date, 6 days).
A native GTX480 should be about 10% faster, having 10% more shaders...
So a native GTX480 on XP could get 77K per day.
On Linux that would be about 81K and with the 11% benefit of CUDA 3.1 (over CUDA 3.0) that would be about 90K per day.

A GTX295 might get that, perhaps even slightly more, but it is unlikely to be as stable and is not as future proofed WRT crunching GPUGrid tasks.

That said, the Fermi's are too expensive.
I would expect that to relax somewhat over the next 2 or 3 months, especially when more variants are released, including the GTX 465 (June 9th is this release date, but dont expect a date for the lesser cards, they may just turn up unannounced):

Suggested GTX465 prices are $250 to $300.
Although the GTX 465 only has 1GB GDDR5 that is enough for here. The 607MHz Core and 1,215MHz Shaders have the same clocks.
The downside is that it only has 352 stream processors; 96 less than the GTX470!
So, it looks more affordable but a poorer performer (27% less than a GTX470). Overall it should be reasonable value, if you want a card that crunches here.

If this really wants to compete with the Radeon 5850, the prices will have to drop a bit more, but that usually comes with time; release prices usually fall.

There are several other suggested Fermis such as the GTS430 with 192 shaders. The expected retail price is between $175 and $200.
The 440 and 450 variants are set to bridge the gap between the GTS430 and the GTX465, with the 450 having 256 shaders.

These cards should rival existing high end cards such as the GTX260 and GTX275 for performance, but not the GTX295 without significant code enhancement.
Anyone with a GTX275 should rethink what sort of card they have; it's no-longer high end, its now mid range!

My advice for anyone thinking about buying a card to crunch with, is to either get a Fermi, a GT 240 or preferably wait for a month (mid June):
There will be more Fermi's to choose from and CUDA 3.1 (+11%) is due for release about that time.
My guess is that within a few months, we will no longer be recommending GT 240's for GPUGrid entry level cards, but perhaps they will hold their own for a while yet.

Writing long speculative ramblings isn't very useful. If you spent the time actually looking at the results various machines are returning you'd be better off. Here's some comparisons for the GTX 295, GTX 480 and GTX 470. ALL are running WinXP

Here's the first GTX 295 I looked at:

http://www.gpugrid.net/results.php?hostid=62542

It averages around 13,800 seconds for a 6,755.61 credit TONI_HERGunb WU. Since it has 2 cores running that's a WU every 6,900 seconds.


Here's by far the fastest GTX 480 that I've found:

http://www.gpugrid.net/results.php?hostid=35174

It averages around 7,800 seconds for a 6,755.61 credit TONI_HERGunb WU.


Here's your GTX 470. It's running faster than any GTX 480 that I've been able to find except the one listed above:

http://www.gpugrid.net/results.php?hostid=71363

It averages around 8,650 seconds for a 6,755.61 credit TONI_HERGunb WU.


For comparison my GTX 260, again running XP, is averaging 12,675 seconds for a 6,755.61 credit TONI_HERGunb WU.

I did spend some time looking for GTX 480 cards. There may be more out there performing as well but I haven't found them.

I did look at many results, especially from top contributors. My speculation was limited to forthcoming GTS400 cards; the GTX465 specs are published.

Going by the best XP GTX295 you could find, running 6.72 (the fastest) tasks it could theoretically get a RAC of (86400/13807)*6755.61=85K:
2363335 1495413 20 May 2010 7:06:09 UTC 20 May 2010 14:50:24 UTC Completed and validated 13,807.71 737.55 4,503.74 6,755.61 Full-atom molecular dynamics v6.72 (cuda)
However that system actually gets a RAC of 63K!

Going by the tasks for Ton’s GTX480:
86400/6416.86*6642.02=90K
2363322 1495382 20 May 2010 7:53:42 UTC 20 May 2010 13:13:37 UTC Completed and validated 6,416.86 7,896.70 4,428.01 6,642.02 Full-atom molecular dynamics v6.73 (cuda30)

Unfortunately it is still too early to look at system RACs for Fermis as the systems need to be up for quite some time to equilibrate.

My GTX470 is reasonably well optimized, but nothing special. I have also recently been having system RAM issues; I only have 2GB in that system and I was running 7 CPU threads on a native i7-920 and some CPU tasks were using over 400MB each. My page file is over 5GB, but it is a testing system!
I will eventually get more RAM and I have just upped the CPU to 3.5GHz, so the times should eventually fall to below 8K sec.

Again, the card is only clocked to the same core speed as a GTX480 and it does not yet use CUDA 3.1, so there is more to come.
Lots of people are struggling with the poor Win7 drivers.
Give me a GTX480 and I could easily get 110K per day!

Anyway, wait a month and see how things pan out before getting a Fermi; there will be a GTX465 version and possibly/hopefully a 512 shader Fermi.

In my opinion, as an experienced buyer and seller of IT equipment, if you have a GTX275, GTX280, GTX285 or GTX295, SELL IT NOW and get a FERMI next month.
Within a month your GTX200 cards will be worth much less!

However that system actually gets a RAC of 63K!

Determining how well a card produces or not is difficult. The average time for a 6.72 HERG were used for the theorical calculations on the best 295 yet the note quoted above does not mention that the card is still running many 6.03 tasks.

For example, my 295 is not returning each WU quite so fast but because I switched over to mostly 6.72 two days ago my RAC is over 76K and climbing.
http://www.gpugrid.net/results.php?hostid=56900

---

Thanks - Steve

However that system actually gets a RAC of 63K!

Determining how well a card produces or not is difficult. The average time for a 6.72 HERG were used for the theorical calculations on the best 295 yet the note quoted above does not mention that the card is still running many 6.03 tasks.

For example, my 295 is not returning each WU quite so fast but because I switched over to mostly 6.72 two days ago my RAC is over 76K and climbing.
http://www.gpugrid.net/results.php?hostid=56900

Exactly, on v6.72 tasks the above GTX 295 is outproducing any GTX 480 to date. SK babbles on about future potential RAC but it's really very easy to determine which card is faster. For some reason he's just not getting it, and it's getting a little frustrating.