cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

bernick
Journeyman III

Best card for opencl GPU computing

What would be the best (e.g. fastest) card to run pure opencl programs (I am not interested in video rendering, only computation power)?

I am thinking about 295x2, 290x or the w9100.

Any recommendations?

cheers.

0 Likes
1 Solution
jason
Adept III

It depends.  For most workloads the 295x2 - count the compute units, look at memory bandwidths.  Those are the easy differentiators within a specific graphics card vendor - its not as straight forward between vendors.  If you want to get particular about it you also have to keep in mind architecture too which the current GCN doesn't do as hot on double floating point (hoping this is fixed in near future)...but its got oodles of memory bandwidth and float/integer stuff is all perfect so as long as doubles aren't what you're after you're safe.  The w9100 is hawaii (pre GCN) so it doesn't have as much memory bandwidth or float/integer math but it's darn competitive on double floating math - it's not every day you go backwards in performance on something as important as doubles so harshly but for some reason they chose to.  The 7970 is also pre-GCN and similiarly does quite good on double floating point math.

The most non-bs way to do this comparison for computational capabilities of card is actually the datasheet so to speak, as long as you can read it.  Best place:

List of AMD graphics processing units - Wikipedia, the free encyclopedia

View solution in original post

0 Likes
3 Replies
jason
Adept III

It depends.  For most workloads the 295x2 - count the compute units, look at memory bandwidths.  Those are the easy differentiators within a specific graphics card vendor - its not as straight forward between vendors.  If you want to get particular about it you also have to keep in mind architecture too which the current GCN doesn't do as hot on double floating point (hoping this is fixed in near future)...but its got oodles of memory bandwidth and float/integer stuff is all perfect so as long as doubles aren't what you're after you're safe.  The w9100 is hawaii (pre GCN) so it doesn't have as much memory bandwidth or float/integer math but it's darn competitive on double floating math - it's not every day you go backwards in performance on something as important as doubles so harshly but for some reason they chose to.  The 7970 is also pre-GCN and similiarly does quite good on double floating point math.

The most non-bs way to do this comparison for computational capabilities of card is actually the datasheet so to speak, as long as you can read it.  Best place:

List of AMD graphics processing units - Wikipedia, the free encyclopedia

0 Likes


jason wrote:



...  The w9100 is hawaii (pre GCN) ...  The 7970 is hawaii also and similiarly does quite good on double floating point math.



The most non-bs way to do this comparison for computational capabilities is card is actually the datasheet so to speak, as long as you can read it.  Best place:



List of AMD graphics processing units - Wikipedia, the free encyclopedia



According to the WP:EN article, w9100 is Hawaii XT while 7970 is some version of Tahiti.

See attached image for a list of architectures pulled out of CodeXL.

0 Likes
maxdz8
Elite

For development or production?

290 and 290x models vary a lot by manufacturer. Some run so hot they're basically louder 280x.

I honestly wouldn't use them for development albeit for production I might consider the Sapphire 290x which has a very solid cooling performance.

Nonetheless, the price premium on them I can get is higher than basically any benefit I could have in density so unless you have some solid offer I'd be careful with those.

As I don't use double float I have never found much interest in FirePro products albeit they're very good.