Archives Discussions

[fire64]

Is there any plausible reason why double precision can only be used on highlevel (and often more expensive) cards?

Where does the restriction comes from? From the driver, the hardware or the ICD-model file?

Is there any hope for double precision support for a HD5670 card in the (near) future - perhaps the next release of ATI-StreamSDK?

Thanks for your explanations!!!

---

PS: I want to avoid to buy a new card like HD58xx - because i would need a new tower, fans and powersupply as well. That's too much for a short test on what i would gain for my scientific project.

[MicahVillmow]

fire64,
Double precision is available only if the hardware has the feature. Double precision hardware is expensive, so only the high end chips have the hardware builtin. The low/middle end chips do not have the hardware, so will not support double precision.

[fire64]

MicahVillmow,

thanks for your short and clear explanation. In contrast to your arguments the following documentation, which was one reason for me to buy that cheap card, shows on page 5 that double precision is (or would be) supported on HD5670. Please correct that for the next release.

http://developer.amd.com/gpu/ATIStreamSDK/assets/ATI_Stream_SDK_Getting_Started_Guide_v2.2.pdf

For my test i will switch to single precision using the cheap but fast GPU and after a few optimization cycles i will switch back to double precision using the CPU. I am not sure this works (roundoff errors significant?), but let's try.

Thanks.

[MicahVillmow]

fire64,
Thanks for pointing that out, I've let our documentation guy know so that he can fix the issue.