1

I've seen this mentioned in various places. From Wikipedia:

For equivalent single precision output, Fermi-based nVidia Geforce cards have four times less dual-precision performance [than Tesla cards].

I've heard vague explanations involving a 'multiplier.' Does it involve something physically disabled in the equivalent Geforce dies?

1 Answer 1

3

It's the difference between having dual-precision operations implemented in hardware and having to do a dual-precision operation as four single-precision operations.

For example, consider multiplying "16" by "43". If you had a native dual-digit multiplier, you would do it in one operation. If you only had a single-digit multiplier, you'd need to multiply 1 by 4, 1 by 3, 6 by 4, and 6 by 3, to do it. Like this:

   16
 X 43
 ----
   18 (6*3)
   3  (3*1)
  24  (4*6)
  4   (4*1)
-----
  688

Notice that you needed to do 4 single-digit multiplications to do one two-digit multiplication, so it takes four times as long. So, if you don't have a dual-precision multiplier but only have a single-one, it takes you four times as long to do a double-precision multiplication.

4
  • what about the cost of the additions? how does that compare?
    – GreyCloud
    Feb 6, 2012 at 10:56
  • Think about it, "34 + 17" requires you to do "4+7" and "3+1", then you have to add the result of those two additions. Then you have to add the carry. So four single-digit additions to do one doule-digit addition. Feb 6, 2012 at 10:58
  • This makes sense. So can each core of a Tesla GPU address 64 bits in one clock cycle, and each core of a Geforce GPU address 32 bits? Feb 6, 2012 at 22:56
  • 1
    It has nothing to do with addressing. It has to do with the width of the arithmetic units. Feb 6, 2012 at 23:12

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .