Is the AMD 7970 graphics card still the best option for PhotoScan?

[Wishgranter]

The "problem" is that OpenCL is still not what it can be ( interesting 1.2 revision with C++ integration and other stuff ), but the goal - independent cross platform solution. And CUDA is  closed enviroment, so it can run just on Nvidia hw.... in say 1-2 years if al OpenCl stuff is implemented and the HSA is in full beauty on market we will see interesting and fast solution. But for now AMD lead the way - raw performance

Have heard that Nvidia will be sold to Intel in 5 years.... Yes the moneymaking stuff, but performance is never enought, they think we will drop to buy new - faster cards, but im think opposite.... But even GPUs are from development perspective near its limits, so every new architecture will not bring 100 % speedup as in last years, but just 20-30 % speedup. therefore they cap the performance. Becasue why you would spend 300-1000 EUR/USD for new card if it have just 10-30 speedup every year    

[David Cockey]

I have a new HP laptop with an Intel I7-4700MQ processor and NVIDIA GeForce GT 740M graphics. Graphics driver is the latest from NVIDIA. Open Hardware Monitor shows very heavy use of the GPU during Build Dense Cloud with PhotoScan 1.0 Build 1748, and the log shows the GPU contributing as much as the CPU. Clearly the GPU is not "crippled".

 

[Wishgranter]

Hi david, the GPU usage is NOT a clear hint on proper and FULL usage.. they can use wait cycles and the processor IS working on 100% but its work is WAIT say few cycles...... So its not the best way to say how well its perform.... and to get the "CLEAN - RAW" results of monitoring you need totally other tols for that.... some developers toll that has acces to info that is accesible to people that can it understand.... AS we have already discussed it wth CPUs, even my dual CPUS are on 100 % but they are 3,5 times slower nder win as under Linux... the problem is, short explained in compilers...... so my CPU is working very hard on 100% but in reallity its make a lot of wait cycles.....

[Marcel]

Results for Crossfire AMD R290 (non-X model):

finished depth reconstruction in 2699.58 seconds
Device 1 performance: 114.716 million samples/sec (CPU)
Device 2 performance: 621.508 million samples/sec (Hawaii)
Device 3 performance: 614.536 million samples/sec (Hawaii)
Total performance: 1350.76 million samples/sec

So not really faster than the 7970, which is a bit of a disappointment. One of the cards seemed to be throttled down 50% of the time according to GPU-Z, but I am not sure how accurate the reporting is. Temperature and fans speeds are both below the limits, so it's not because it's overheating. This was using 8 out of 12 CPU cores (I hoped freeing up extra cores would keep the GPUs busy).

By the way, is there a specific test file you guys use to measure the performance?

[Wishgranter]

Marcel try use scene from this thread http://www.agisoft.ru/forum/index.php?topic=651.0

disable ALL CPUs in the OpenCL settings and run the dataset......

Have read that 290X are not the best Compute cards, but it need to be tested on same dataset to be sure.... 

[Marcel]

Here are the results for the 'official' benchmark file, point cloud on Ultra:

finished depth reconstruction in 133.39 seconds
Device 1 performance: 765.774 million samples/sec (Hawaii)
Device 2 performance: 762.169 million samples/sec (Hawaii)
Total performance: 1527.94 million samples/sec

So with the official file, the R290 is about 25% faster than a 7970.

[Wishgranter]

Thanx Marcel..... then its inside expectations...

[Exhale]

Here are the results for the 'official' benchmark file, point cloud on Ultra:

finished depth reconstruction in 133.39 seconds
Device 1 performance: 765.774 million samples/sec (Hawaii)
Device 2 performance: 762.169 million samples/sec (Hawaii)
Total performance: 1527.94 million samples/sec

So with the official file, the R290 is about 25% faster than a 7970.

Hi Marcel,

Is this result for one R290 card ?

Guys,  I wanted to get 7970, however it is out of stock.. What else do you recommend ? R290x or Nvidia series :=)

Cheers

[Marcel]

Hi Marcel,

Is this result for one R290 card ?

Guys,  I wanted to get 7970, however it is out of stock.. What else do you recommend ? R290x or Nvidia series :=)

Cheers

No, each card is doing between 700 and 900 million samples per second (it depends on the project).

The 7970 is still gives the best performance for your money. The R9 290 is slightly faster but more expensive (and runs a lot hotter as well). If the 7970 is out of stock you can consider the AMD R9 280X, which is the re-branded version of the 7970.

nVidia cards are still not so good for Photoscan. I tested a GTX760 and it was giving around 450 million samples a second. The Titan has similar performance as the R9 290, but at double the price.

In practice if you have 2x 7970 you'll be fine. Don't focus too much on the GPUs, you'll need a fast CPU as well for all the phases that are not GPU accellerated.

[Exhale]

Hi Marcel,

Thanks for your reply,
CPU will be  INTEL i7 4930K 3.4 GHz Socket 2011 12MB 6 Core, I believe it will do.
Thanks for the  R9 280X recommendation. I'll focus on it. The Last and expensive option will be R9 290 series. 

Thx

[Marcel]

I have the same CPU, it's very fast. Keep in mind that this CPU is limited to a maximum of 64GB memory. For very big projects you will need more memory for the Mesh generation stage (or you have to build in separate chunks).

Unfortunately the only other option (Xeons) makes things a lot more expensive.

[Subunderground]

Our rig with 2 AMD 7970's is about 20% to 30% quicker (at GPU stage) than our identical rig with a Titan. Not very scientific, but cheaper 

[Exhale]

Thx for the information.
I'll go for Sapphire Radeon TOXIC R9 280X 3GB.
I think, I will be able to post the result this weekend.   

[blmvnew]

Hi All,

I am using 7990.
The result of each 7970 core is between 650~950 million samples per second.
So the total result of one 7990 is 1300~1900 million samples per second.
Think about the cost for 7990, it is realy the best one.

If you have more power from the PS and one more PCIE slot.
Two 7990 in the system will be the super machine for photoscan processing.

Blmvnew

[Marcel]

From some reviews it seems that two 7990 cards in one machine will cause them to overheat.  It might work,  but you'll need really good case cooling.