Sign in with
Sign up | Sign in
Your question

Tesla VS A Bunch of 480s

Last response: in Graphics & Displays
Share
May 2, 2010 10:16:55 PM

Greetings all, i am computer science undergrad and currently working on a project that involves testing the efficiency of parallel code using CUDA. Budget is of no issue, and i am currently in the process of picking out the desired GPU to run the experiments on. Yet, i seriously cannot decide on whether to get a SINGLE Fermi Tesla or 4 Geforce 480s.
I know that Tesla is specifically built in order to perform parallel computations, but wouldn't 4 480s outperform a single Fermi Tesla?

More about : tesla bunch 480s

a b U Graphics card
May 2, 2010 10:50:27 PM

It might, but remember that the 480's have relatively crippled chips for computing, infact their doubleprecision capability is only 25% of their design. The Tesla might be a better option if your doing straight computing, as it is optomized for that. The GTX480s will work, but like using Quadros for gaming, will not do as well at it as something designed for it.
a c 173 U Graphics card
May 2, 2010 10:59:41 PM

Tesla boards are only for computing and lack ports so they do not display while a GTX480 do. If you buy the Tesla it's use is very limited out side computing so keep that in mind when you make a choice.

Related resources
a b U Graphics card
May 2, 2010 11:01:12 PM

The 4 GTX 480's outperform a single TESLA on many fields namely heat (enough to melt the arctic) and noise (enough to make you a monk if you can train your brain to concentrate through that turbulence).. Even if it does outperform the TESLA unit for parallel computing, it is not worth it.. Also, for that price you pay, you are not going to get the support that you'd enjoy with TESLA..
May 2, 2010 11:09:02 PM

Alright, thank you for your responses. I think i might go for a Tesla after all, and yes i do know that Tesla do not output video.
Damn, now my mind is getting boggled over the possibility of being able to rig a water cooling system unto the Tesla so i can increase core frequency :p .
a b U Graphics card
May 2, 2010 11:19:41 PM

What for do you need to increase the core frequencies..?? They are too expensive cards to loose the warranty on..
May 2, 2010 11:22:05 PM

I live in Lebanon, warranty won't apply here. I will have to ship the Tesla from the states.
a b U Graphics card
May 2, 2010 11:25:59 PM

Emperus said:
What for do you need to increase the core frequencies..?? They are too expensive cards to loose the warranty on..


Agreed, it's the same reson you don't overclock Quadros. If you need more power, just buy another one if budget is no issue.
a b U Graphics card
May 2, 2010 11:41:53 PM

frisbeetarian said:
I live in Lebanon, warranty won't apply here. I will have to ship the Tesla from the states.


The point is that you might destroy the cards.. And since you mentioned water cooling, i guess your plans are for a 30% or higher overclock which, not might but will surely cause trouble.. These cards are meant to run at those clocks and if extra boost in clock speeds would've been causing any advantages then Nvidia would surely have done it themselves..
a b U Graphics card
May 2, 2010 11:54:05 PM

The Tesla cards are not like normal GPUs as stated above, they are not designed to be overclocked, their design is built to a certain tolereance, and any overclocking will put them out of that range. The biggest problem with overclocking the Teslas is the power draw....a Tesla already requires to run WELL a 750W PSU, and any way of overclock will increase power draw requirement.
May 2, 2010 11:56:09 PM

Emperus said:
The point is that you might destroy the cards.. And since you mentioned water cooling, i guess your plans are for a 30% or higher overclock which, not might but will surely cause trouble.. These cards are meant to run at those clocks and if extra boost in clock speeds would've been causing any advantages then Nvidia would surely have done it themselves..


They settle on, what is it, 1.3ghz per core on the Fermi Tesla because the stock cooling solution can only be effective at these frequencies. Custom rigging a water cooling solution might be able to net me an improvement in the margin of 10%. Im not planning to go beyond that, i just know that if i provide the proper cooling, the frequency of the cores can effectively be upped. I wont go overboard though, i think a max increase of 8% should be fine and not damage anything.
May 2, 2010 11:57:38 PM

sk1939 said:
The Tesla cards are not like normal GPUs as stated above, they are not designed to be overclocked, their design is built to a certain tolereance, and any overclocking will put them out of that range. The biggest problem with overclocking the Teslas is the power draw....a Tesla already requires to run WELL a 750W PSU, and any way of overclock will increase power draw requirement.


Can you please explain why they are not designed to be overclocked?
a b U Graphics card
May 3, 2010 12:12:31 AM

frisbeetarian said:
Can you please explain why they are not designed to be overclocked?


The Tesla cards, when they were first designed are designed for supercomputing applications, namely float point operatins, which is why CUDA came about. Basically you could have a small cluster of powerful cards that would do the same if not more, compared to a cluster of Linux machines. While more expensive a Tesla cluster is smaller, although more expensive.

Since they were designed for supercomputing applicaitons, they were built to certain tolereances that are designated peaks of performance, and clocked accordingly. While yes you could overclock them, the problem is that it would likely lead to instability as they were built to be run as is, not to be overclocked. This holds doubly true if you are running more than one Tesla card. I firmly believe that the PCB and architecture should be able to handly an overclock, but while it may be able to handle it, that dosen't mean it would be stable doing so.
a c 173 U Graphics card
May 3, 2010 12:57:04 AM

The tolerances argument does hold water but there is perhaps one more issue on top and that is software. I am doubtful that there is any utilities out there that allow overclocking of Tesla and more so if Nvidia disabled the feature.
a b U Graphics card
May 3, 2010 12:59:05 AM

Overclocking can't possibly damage a card.

Now overvolting can, that can go wrong so fast. The question is about safe overvolting. I believe the fermis are designed for lower than the current stock voltage but because their manufacturing process was so bad the transistors didn't come out so good. To get 512 working CUDA cores they upped the voltage quite a bit and with that heat skyrocketed. So even with water, you might not want to further overvolt the cards.


But if anyone has water on a fermi. Definitely get as far as possible on stock voltages.
a b U Graphics card
May 3, 2010 1:41:06 AM

rofl_my_waffle said:
Overclocking can't possibly damage a card.

Now overvolting can, that can go wrong so fast. The question is about safe overvolting. I believe the fermis are designed for lower than the current stock voltage but because their manufacturing process was so bad the transistors didn't come out so good. To get 512 working CUDA cores they upped the voltage quite a bit and with that heat skyrocketed. So even with water, you might not want to overvolt the cards like crazy.


But if anyone has water on a fermi. Definitely get as far as possible on stock voltages.


Thats true, and given the fact that NVidia has indeed disabled any way to overclock it other than increasing the PCIe voltage, it creates even more of a chance to fail. There are currently no utilities to overclock a Tesla card, and overvolting is dangerous.

This is not to mention the fact that I'm not sure any manufacture makes a water cooling solution for Tesla cards.

After all this, if you still want to do it, +.2V is the absolute max. I would leave it alone personally, like I said, if budget is no limit, get a second Tesla card, it would completely elimate any performace gains you would have by overclocking/overvolting.
May 3, 2010 2:17:04 AM

While I'm a gamer, and I would much rather own 4 gtx 480's than a single fermi telsa card, with similar cuda performance. But, I think if I was gonna use the card for scientific research, I would definitely go with the telsa card. Running four gtx cards is gonna be expensive.

There is no way in hell I would run quad sli gtx 480's without a water cooling set up. Then you have to take into account that in order to run quad sli, you would need a 1500watt PSU, a high end water cooling setup, 4 water blocks, a big ass case, a quad sli m.b., not to mention the electric bill of running all that. If your rich, and you want to game, go with the quad sli, but if your not a gamer, then I would go with the single tesla card.

Although, you could consider compromising and just go sli with two gtx 480's. yeah, it might not be as fast a the fermi tesla, but at least you can game with it.

Another key point to remember is fermi tesla can run much quicker then the older tesla cards, and do double precision, so if your looking to upgrade, even going with a dual gtx 480 system should be much quicker then going with the last generation tesla cards.
a b U Graphics card
May 3, 2010 2:43:28 AM

dreamphantom_1977 said:
While I'm a gamer, and I would much rather own 4 gtx 480's than a single fermi telsa card, with similar cuda performance. But, I think if I was gonna use the card for scientific research, I would definitely go with the telsa card. Running four gtx cards is gonna be expensive.

There is no way in hell I would run quad sli gtx 480's without a water cooling set up. Then you have to take into account that in order to run quad sli, you would need a 1500watt PSU, a high end water cooling setup, 4 water blocks, a big ass case, a quad sli m.b., not to mention the electric bill of running all that. If your rich, and you want to game, go with the quad sli, but if your not a gamer, then I would go with the single tesla card.

Although, you could consider compromising and just go sli with two gtx 480's. yeah, it might not be as fast a the fermi tesla, but at least you can game with it.

Another key point to remember is fermi tesla can run much quicker then the older tesla cards, and do double precision, so if your looking to upgrade, even going with a dual gtx 480 system should be much quicker then going with the last generation tesla cards.


True, but given he's trying to "testing the efficiency of parallel code using CUDA" I would still stick with the Fermi, especially if it's on a research machine/workstation.
May 7, 2010 4:36:21 PM

Thank you all for the answers provided, i will take into consideration whether it would pay off to overcloak or not. Another question, the Teslas work on the regular PCI Express slot right?
a c 173 U Graphics card
May 7, 2010 5:02:50 PM

frisbeetarian said:
Thank you all for the answers provided, i will take into consideration whether it would pay off to overcloak or not. Another question, the Teslas work on the regular PCI Express slot right?



It should and same for power. These are often bought by some for encoding as well so you should be good.
a b U Graphics card
May 7, 2010 10:25:00 PM

1. The Fermi cards have only 25% of the Double Precision cores compared to the Tesla cards. If your cose,etc uses these cores you are probably better off with the Tesla, else get the Fermi.

2. The drivers for Tesla are optimized for those specific areas (scientific computing,etc)

3. Overclocking DOSE NOT void your warranty AS LONG AS you DO NOT tell them you have Overclocked. ONLY BIOS modding,etc voids your warranty. nVidia,etc have no way of knowing if you have overclocked the card.

4. Also, DO realize that running 4x GTX480s in SLI,etc will require a good water cooling set up as there isn't much space between the cards with air cooling solutions.
!