cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

rahulgarg
Adept II

Kaveri APUs and fp64 GPU support?

Hi. I was wondering whether the GCN GPU in Kaveri APUs such as 7850K supports fp64 and if so, at what speed?

0 Likes
10 Replies
pinform
Staff

Yes, Double Precision Floating Point Operations (FP64) are SUPPORTED.
You can run double-precision floating point operations at full, half or quarter rate compared to the chip’s native maximum rate.

0 Likes

>You can run double-precision floating point operations at full, half or quarter rate compared to the chip’s native maximum rate.

How do you set this?

As far as I can tell its 1:16 only

0 Likes


pinform wrote:



Yes, Double Precision Floating Point Operations (FP64) are SUPPORTED.
You can run double-precision floating point operations at full, half or quarter rate compared to the chip’s native maximum rate.


I think we need a little clarification on this point.  What is the A10-7850K's native maximum rate for DP FLOPS for the GPU cores?

0 Likes

An updated documentation would help on this. I hope we won't have to wait till summer for it.

0 Likes


moozoo wrote:



AnandTech Portal | Floating point peak performance of Kaveri and other recent AMD and Intel chips


I would like to point out that the author of article, Rahul Garg, is also the original poster in this thread.  Rahul may be basing his numbers on moozoo's 1:16 guess.  This is why we need good documentation from AMD.  Or bad/unreliable information just gets copied and repeated...

0 Likes

See

http://forum.beyond3d.com/showpost.php?p=1820427&postcount=1103

739.7/46.27 = 15.99

If the chip could do better than 1:16 then they would have contacted Rahul and had the article corrected.

"You can run double-precision floating point operations at full, half or quarter rate compared to the chip’s native maximum rate."

For all other AMD GPU's this is hardwired based on chipset model. It isn't adjustable via software.

It would be great if they made it adjustable like on Nvidia's Titan.

0 Likes


moozoo wrote:



See


http://forum.beyond3d.com/showpost.php?p=1820427&postcount=1103



739.7/46.27 = 15.99



If the chip could do better than 1:16 then they would have contacted Rahul and had the article corrected.



"You can run double-precision floating point operations at full, half or quarter rate compared to the chip’s native maximum rate."


For all other AMD GPU's this is hardwired based on chipset model. It isn't adjustable via software.


It would be great if they made it adjustable like on Nvidia's Titan.


So it appears that it is a 1:16 ratio, but this is based on a benchmark result.  However unlikely, it's conceivable that the compiler use didn't generate the optimal fp64 instructions for whatever the benchmark is running.  It's likely spinning on multiply-add operations and a broken compiler could split it into two operations even if there were a fp64 MAD instruction available on the hardware.

I would still like to see some documentation from AMD on this chip...

0 Likes

Hi. Just to clarify, the data reported in the anandtech article is not from this thread. We were able to test 7850K internally using OpenCL and also we were able to get an official response from AMD's PR rep over email that it is indeed 1:16.

0 Likes
moozoo
Adept III

I'm hoping these have a better fp ratio.....

AMD1311.1 = "KV SPECTRE WORKSTATION 100W (1311)"

AMD1310.1 = "KV SPECTRE WORKSTATION 65W (1310)"

from the spoiler of the first post here AMD Catalyst 13.x (13.150.0.0 May 23) - Guru3D.com Forums

0 Likes