Comment
from @ 2010/02/05
The first CUDA cards introduced a while ago also sucked at DP, so Tesla cards aren't the reason why the 'older' cards don't do DP very well.
DP is indeed interesting for scientific groups who want accuracy, but sometimes real-time processing is more important that a bit more accurate results.
Quote:
As most scientific computing groups consider accuracy to be paramount most of them use double-precision workloads. It was specifically done to ensure groups couldn't build FASTRA computers without paying more for Tesla cards. They'd get better performance from RV870 instead of GeForce now.
|
DP is indeed interesting for scientific groups who want accuracy, but sometimes real-time processing is more important that a bit more accurate results.
Comment
from @ 2010/02/04
As most scientific computing groups consider accuracy to be paramount most of them use double-precision workloads. It was specifically done to ensure groups couldn't build FASTRA computers without paying more for Tesla cards. They'd get better performance from RV870 instead of GeForce now.
Comment
from @ 2010/02/04
DP is useful because it delivers more accurate results
Quote:
Refers to a type of floating-point number that has more precision (that is, more digits to the right of the decimal point) than a single-precision number. The term double precision is something of a misnomer because the precision is not really double. The word double derives from the fact that a double-precision number uses twice as many bits as a regular floating-point number. For example, if a single-precision number requires 32 bits, its double-precision counterpart will be 64 bits long. The extra bits increase not only the precision but also the range of magnitudes that can be represented. The exact amount by which the precision and range of magnitudes are increased depends on what format the program is using to represent floating-point values. |
Comment
from @ 2010/02/04
Then don't use DP
Comment
from @ 2010/02/03
Won't be the case anymore... double-precision performance has been limited to 1/4th performance that is capable on GeForce cards... have to buy Tesla models for full performance. RV870 will perform better than GeForce models because of this in DP work.
Comment
from @ 2010/02/03
Actually, Tegra is the mobile name ... not terga
Comment
from @ 2010/02/03
OK, so they can stop laughing now. Tones considered Tesla's for FASTRA II, but it was 5x as expensive for the same performance.
Comment
from @ 2010/02/03
Comment
from @ 2010/02/03
The long wait
Comment
from @ 2010/02/03
And its price ?
Comment
from @ 2010/02/03
this is a picture of NVIDIA Tesla product line :-)
Comment
from @ 2010/02/03
More facts ?
DP is indeed interesting for scientific groups who want accuracy, but sometimes real-time processing is more important that a bit more accurate results.
DP FMA rate:
GT200 offers 88.5 Gflops
GF100 Tesla offers 744 Gflops
GF100 GeForce offers 186 Gflops
(Source TheTechReport)
I'm not sure what percentage of the market writes for SP FMA and runs calculations twice versus DP FMA and runs them once, but there are enough organizations that built GT200 GeForce computing platforms that this will certainly affect quite a few of them.