Double Precision Flops per Cuda or Stream, Kepler vs Tahiti vs Femi

Can someone pass along some baseline data relating double precision performance of the various architectures.

Double Precision Flops per Cuda or Stream, Kepler vs Tahiti vs Femi. It would be great to have this per MHz and per Watt.

It appears that Nvidia hides this now. Even a URL to Nvidia or AMD that lists double-precision numbers would be great.

2 answers Last reply
More about double precision flops cuda stream kepler tahiti femi
  1. At the moment that's a little difficult. The problem is that Nvidia have since Fermi had cards which are designed more as gaming chips and cards which are designed to be both gaming but heavily siding with compute performance chips.
    The GK104 is the gaming performance chip and as such its DP is artificially capped reduced to allow for more gaming performance.
    AMD have released the 7970 which has more compute performance compared to the cards it replaces.

    So while these numbers are available I don't really see what comparing the cards on that basis would be worth.

    Mactronix :)
  2. A lot also depends on the software used as some software requires only CUDA, while others (though much fewer) require only AMD's thing (whatever it is called), and yet more software is generally agnostic.

    Unless CUDA is required I would go with the AMD 7970, but if CUDA is required then go with the 570 or 580 because the 680 is junk when it comes to this type of workload. Also, if the budget permits, look into Quadro and FireGL cards as those are made for this type of workload.
Ask a new question

Read More

Graphics Cards Performance Nvidia Graphics