Multiple GPUs with different numbers of SPs
Moderators: NeilBlanchard, Ralf Hutter, sthayashi, Lawrence Lee
Multiple GPUs with different numbers of SPs
In my new rig I originally bought a GeForce 9600GT because it's the most powerful card I could find that was passively cooled. At the time I didn't know that FAH could run on a GPU.
Now I do and I'm excited about it.
A couple of days ago I received an 8800GT I bought on eBay for $85. The purpose of this card was to fold. Period. Nothing else.
When I went looking for specific instruction on how to configure multiple GPUs, I found several posts warning of restrictions when two GPUs with different numbers of stream processors were used. [Sadly I had not seen these posts prior to ordering the new card.] The posts claimed that if the number of stream processors were not even multiples of each other (e.g. 64 and 128), that the card with more SPs would only use the number found on the lesser card. The 8800GT has 112 SPs, the 9600GT 64, so according to these sources, the 8800GT could be expected to use just 64SPs and therefore run no faster than the 9600GT.
I'm JUST starting to look into this but it appears to me that they were partly right, but it may depend on which card is GPU 0 or it may depend on the order the clients start.
The 9600GT alone was doing 3K PPD
The 8800GT alone was doing 5K PPD
With both cards in the system, the 8800GT is still doing 5K PPD but the 9600GT is only doing 2300 PPD. The 8800GT is still running at full speed but the second card, the 9600GT, is running at a lower performance level. It it possible the 9600GT is only using 56 (half of 112) SPs?
Now I do and I'm excited about it.
A couple of days ago I received an 8800GT I bought on eBay for $85. The purpose of this card was to fold. Period. Nothing else.
When I went looking for specific instruction on how to configure multiple GPUs, I found several posts warning of restrictions when two GPUs with different numbers of stream processors were used. [Sadly I had not seen these posts prior to ordering the new card.] The posts claimed that if the number of stream processors were not even multiples of each other (e.g. 64 and 128), that the card with more SPs would only use the number found on the lesser card. The 8800GT has 112 SPs, the 9600GT 64, so according to these sources, the 8800GT could be expected to use just 64SPs and therefore run no faster than the 9600GT.
I'm JUST starting to look into this but it appears to me that they were partly right, but it may depend on which card is GPU 0 or it may depend on the order the clients start.
The 9600GT alone was doing 3K PPD
The 8800GT alone was doing 5K PPD
With both cards in the system, the 8800GT is still doing 5K PPD but the 9600GT is only doing 2300 PPD. The 8800GT is still running at full speed but the second card, the 9600GT, is running at a lower performance level. It it possible the 9600GT is only using 56 (half of 112) SPs?
Here is one good discussion about it, but with no clear conclusions except that mixing cards with different numbers of shaders typically leads to a loss of performance in at least one of the cards vs. running that card alone.
http://foldingforum.org/viewtopic.php?f=52&t=7797
http://foldingforum.org/viewtopic.php?f=52&t=7797
-
- *Lifetime Patron*
- Posts: 4284
- Joined: Fri Apr 04, 2003 6:21 pm
- Location: Undisclosed but sober in US
I'm only saying I won't buy 2 cards that have a different # of processors. The idea of buying a 2nd card and then have one slow because of the other sounds absurd, but there it is.
I've actually complained to EVGA about rebates that apply only to 1 card when folders may buy 2, 3, even 4 cards at a time.
I've actually complained to EVGA about rebates that apply only to 1 card when folders may buy 2, 3, even 4 cards at a time.
Gotcha.
I ordered one of the "new" GeForce GTS 250's this morning. It will replace the 112-shader 8800GT. I'll be able to verify for myself whether the 64-shader 9600GT runs faster when paired with a card with exactly twice the number of shaders vs. the current 64/112 mismatch.
I say "new" because the GTS 250 is basically just a 9800 GTX+ with a new name. There are a few subtle differences.
I ordered one of the "new" GeForce GTS 250's this morning. It will replace the 112-shader 8800GT. I'll be able to verify for myself whether the 64-shader 9600GT runs faster when paired with a card with exactly twice the number of shaders vs. the current 64/112 mismatch.
I say "new" because the GTS 250 is basically just a 9800 GTX+ with a new name. There are a few subtle differences.
-
- *Lifetime Patron*
- Posts: 4284
- Joined: Fri Apr 04, 2003 6:21 pm
- Location: Undisclosed but sober in US
The 1GB version of the GTS250 makes no sense to me. 128 stream processors and 1GB memory for $175, or the GTX260, with 216 stream processors and 896MB of memory.
As proteins get bigger the 260 is going to win by a larger margin.
Good luck with your card, I wish they would have managed to reduce it to a single slot.
As proteins get bigger the 260 is going to win by a larger margin.
Good luck with your card, I wish they would have managed to reduce it to a single slot.
I opted for a 512MB version for $130. The 1 GB models are $150, but yeah, at that point why not step up to the 260?
However, with the 260, you either run it alone or with another 260. No other card would play well with it because of its 216 stream processors.
There's no such thing as a one-slot graphics card. Not with any power anyway. I'll be happy if the 250 take up only two slots. If I have to put an aftermarket cooler on it, like an Arctic Cooling Twin Turbo, it'll take up 3 slots and running a second one may not be an option because there might not be enough room at the bottom of my motherboard.
However, with the 260, you either run it alone or with another 260. No other card would play well with it because of its 216 stream processors.
There's no such thing as a one-slot graphics card. Not with any power anyway. I'll be happy if the 250 take up only two slots. If I have to put an aftermarket cooler on it, like an Arctic Cooling Twin Turbo, it'll take up 3 slots and running a second one may not be an option because there might not be enough room at the bottom of my motherboard.
-
- *Lifetime Patron*
- Posts: 4284
- Joined: Fri Apr 04, 2003 6:21 pm
- Location: Undisclosed but sober in US
You're right about the 260. At 192 processors it was compatible with 96 sp, but all of those are 65nm models, and who wants that?
The most powerful 1 slot solution is the 9800GT, which I can't get very excited about.
The other thing about the 260 is it will work well with the big Thermalright HR03X cooler and say a 120 Yate Loon at 7 volts. I'm assuming if it's folding 24/7 the fan make want to turn on, on occasion. Yuck.
I think the GTX295s get loud. And they are overpriced for what they are, but not as overpriced as the GTX285.
I told EVGA I'd buy a GTX260 today if I knew for a fact that a 40nm compatible model would come out within the 90 day trade-in limit, but all they said was sorry, meaning they can't guarantee when the 40nm stuff will be out. And seeing that NVidia is leaning towards laptop chips first the rest of us may have to wait quite a bit longer.
NVidia dropped the price of the GTX260 today, but not much.
The most powerful 1 slot solution is the 9800GT, which I can't get very excited about.
The other thing about the 260 is it will work well with the big Thermalright HR03X cooler and say a 120 Yate Loon at 7 volts. I'm assuming if it's folding 24/7 the fan make want to turn on, on occasion. Yuck.
I think the GTX295s get loud. And they are overpriced for what they are, but not as overpriced as the GTX285.
I told EVGA I'd buy a GTX260 today if I knew for a fact that a 40nm compatible model would come out within the 90 day trade-in limit, but all they said was sorry, meaning they can't guarantee when the 40nm stuff will be out. And seeing that NVidia is leaning towards laptop chips first the rest of us may have to wait quite a bit longer.
NVidia dropped the price of the GTX260 today, but not much.
I am now folding with a GTS 250 (aka 9800 GTX+) with a second on the way. $137 with shipping from the egg. It's an eVGA card and the OEM fan is running at only 40% on auto and is not audible in my rig.
The PPD are all over the map. 4700 on the current WU.
The PPD on my second card, a 64-shader 9600 GT, is 2857 on the current WU, or about 60% the PPD of the 128-shader card. That's about where it should be. The PPD dropped when paired with a 112-shader 8800 GT, but it's running full-speed when paired with the 128-shader GTS 250.
The PPD are all over the map. 4700 on the current WU.
The PPD on my second card, a 64-shader 9600 GT, is 2857 on the current WU, or about 60% the PPD of the 128-shader card. That's about where it should be. The PPD dropped when paired with a 112-shader 8800 GT, but it's running full-speed when paired with the 128-shader GTS 250.
Welcome to the GPU club. I have a Gigabyte GTX+ that I picked up about 6 or 7 weeks ago on closeout for $140 for the 1gb model. I felt pretty smug about it, only to realize Gigabyte (the brand) uses the slowest memory of any GTX+; only 2000mhz and everybody else uses 2200 or faster, then of course the GTS250 and price drops came along. Oh well, that's the story with computers, there's always something newer, faster, cheaper around the corner.haysdb wrote:The PPD are all over the map. 4700 on the current WU.
As for PPD, on my GTX+ it broke out something like this:
By Point value of WU, the PPD:
353: 6500
384: 6100
420: 5100
511: 4100
1680: 5100
So anywhere from 4100 to 6500 depending on WU. My long term average according to the stats websites is around 4500ppd, but that's with taking some breaks here and there for gaming, as well as full-server backups. A dedicated 24x7 folding machine would probably do closer to 4800 or so with such a card, or maybe more if you overclock it a bit and have the faster memory.