Forum

Author Topic: NVIDIA Tesla Benchmarks  (Read 7333 times)

dtmcnamara

  • Jr. Member
  • **
  • Posts: 73
    • View Profile
NVIDIA Tesla Benchmarks
« on: April 18, 2015, 04:21:56 AM »
So I got tired of ray tracing taking forever and picked up a pair of Tesla cards. Figured I would give them a run at Agisoft and share some informaiton on them:

Nvidia Tesla M2090
6GB DDR5
225W TDP
Average power usage with Agisoft 150-180w
Idle power usage 20-30w
Idle temps 30-47C
Average temps under load 70-90C (passive cooling right now)
ECC Memory State turned ON

I am getting a lot of throttling use to temps right now, the case was not designed for passive cooling co-processors.  I have a few ideas I will be testing out this weekend to try adn get the load temps to under 70C all the time.  Ocne I do I will run the tests again. After multiples tests with the Agisoft benchmark data I got the following average samples/sec

Medium Reconstruction 500-600 million samples/sec

I think I might be able to squeeze a little more out of the cards by turning off ECC memory state, which will raise the memory bandwith. I will test this out tomorrow after I work on getting the cards cooler.

All in all I am very happy with these cards and their performance. They are under permorning the $300 GTX 780 cards I have, but at the price poitn I got them for, $75 each, they are simply amazing.

Just wanted to share since there has been a few other interested in Tesla/Phi chard performance. Check back later this weekend for more results
« Last Edit: April 18, 2015, 02:54:12 PM by dtmcnamara »

Wishgranter

  • Hero Member
  • *****
  • Posts: 1202
    • View Profile
    • Museum of Historic Buildings
Re: NVIDIA Tesla Benchmarks
« Reply #1 on: April 18, 2015, 06:11:16 PM »
with ecc disabled you should get 10-15% speedup = for best performance and from experience  a "gamer" editions of Nvidia cards are much faster an cheaper. The Tesla and Quadro card are targeted to different task, mostly they  pro is use DoublePrecision = DP in scientific calculations

for same price can get 2-3 gamer card that are much much faster...   
----------------
www.mhb.sk

dtmcnamara

  • Jr. Member
  • **
  • Posts: 73
    • View Profile
Re: NVIDIA Tesla Benchmarks
« Reply #2 on: April 18, 2015, 08:52:42 PM »
for same price can get 2-3 gamer card that are much much faster...   

These were $75

Wishgranter

  • Hero Member
  • *****
  • Posts: 1202
    • View Profile
    • Museum of Historic Buildings
Re: NVIDIA Tesla Benchmarks
« Reply #3 on: April 19, 2015, 11:11:45 AM »
hmmm nice price then...  8)
----------------
www.mhb.sk

dtmcnamara

  • Jr. Member
  • **
  • Posts: 73
    • View Profile
Re: NVIDIA Tesla Benchmarks
« Reply #4 on: April 20, 2015, 04:28:10 PM »
Well these cards run HOTTTTTTTT! I thought my setup with a few 10,000RPM server fans forcing air into the cards would be good enough but I was wrong. I created a funnel to push air into each card and was able to do some basic 2-3 minute tests, but in those short periods the cards would ramp from 30-40C to almost 100C so anything over a small 50 picture project all at medium settings is not possible at the moment.

I ordered some water block for the cards on Sunday so I should be able to get some good tests not this coming weekend but the following weekend. I will also do a write up on cooling these cards just in case someone else goes this route down the road and needs to cool them as well.

Marcel

  • Sr. Member
  • ****
  • Posts: 309
    • View Profile
Re: NVIDIA Tesla Benchmarks
« Reply #5 on: April 20, 2015, 08:57:32 PM »
Are you using server fans? Because these fans are small but displace a lot of air (and are LOUD, another reason not to use this case XD)

That's strange, because that is how these kind of cards are often used:

http://www.beyond3d.com/images/articles/tesla/S870_34_B-big.jpg

http://www.nvidia.com/docs/IO/43395/NV_DS_Tesla_S1070_US_Jun08_NV_LR_Final.pdf

You can see the row of fans in the front. This is an nVidia Tesla S1070 case, you can pick them up for cheap on eBay. The included M1060 GPUs are not  good for Photoscan (around 350 million samples/sec or something).

You can't put consumer cards in it because of the mounting pins (and using 4 cards is not very effective anyway). Not sure if your Tesla cards would fit either since they are newer models.

dtmcnamara

  • Jr. Member
  • **
  • Posts: 73
    • View Profile
Re: NVIDIA Tesla Benchmarks
« Reply #6 on: April 21, 2015, 02:42:42 PM »
The fans im using were pulled from a 4U server. The reason they are running how is just simply becuase im not getting enough air directed through the cards, its going over and under....ill just wait for the water blocks to arrive, should help cool the room down as well.