Archives Discussions

[kugeleck]

OpenCL with a 64bit real-time Linux kernel

Hi.

NVIDIA limits the double precision performance of their consumer graphics cards to 25% of the possible performance. The GeForce GTX 550 Ti has 691.2 GFLOPs [1] and therefore would be capable of 345.6G double precision floating point operations per second. It is limited to 86.4G though. So for Fermi-based consumer cards the double precision performance is 1/8 of the single precision performance.

Does AMD also limit the double precision performance?

According to what I red double precision performance is 1/5 of single precision performance with AMD GPUs. This means that the Radeon HD 5850 has a double precision performance of 417.6G which is 20% of 2088 GFLOPs [2] .

I want to use OpenCL with a 64bit Linux real-time kernel. The NVIDIA driver supports this but as far as I know the AMD one does not? Is real-time support planned? I found information online on how to use the AMD driver in a real-time environment. That requires quite a few code changes though. Then again with the AMD card I would get almost a 400% performance increase.

Taking all of the above into account would it be better to buy a Radeon HD 5850 for OpenCL programming since it costs the same and also supports OpenCL1.1?

[1] http://en.wikipedia.org/wiki/GeForce_500_Series#GeForce_500_Series

[2] http://en.wikipedia.org/wiki/Comparison_of_ATI_graphics_processing_units#PCIe_.28HD_5xxx.29

[dragonxi4amd]

Hi,

Is it really so that developers only get consumer graphics cards and

therefore only 25% of the possible performance -

instead of developer card with 25 % price and 200% performance!?

[kugeleck]

Is it really so that developers only get consumer graphics cards and therefore only 25% of the possible performance - instead of developer card with 25 % price and 200% performance!?

I don't really get your question. I could of course get a FirePro or Tesla card but these cost a lot more money (4-5 times as much for the same performance).

Take these 3 pretty similar AMD cards (number of cores, frequencies, performance, power consumption, ...) for example:

• ATI FirePro V7800 (2GB) with 403.2 GFLOPs (double precision) costs 450€
• SAPPHIRE Radeon HD 5850 (2GB) Toxic with >417.6G GFLOPs (double precision) costs 175€
• SAPPHIRE Radeon HD 5850 (1GB) Xtreme with 417.6G GFLOPs (double precision) costs 100€

Since I would use the graphics card only for OpenCL computation without a display even attached to it I don't see any reason for why I should spend more money for less.

[himanshu_gautam]

The double performance is by design. The VLIW units have to work together so the max GLOPS are reduced. 5850 is a nice card. although you can also look in 6000 series.

[kugeleck]

Only the Radeon HD 6950 and up of the 6000 series support double precision [1] which is very important to me. Those cards also draw to much power for my taste (ca. 150W is my limit).

[1] http://en.wikipedia.org/wiki/Comparison_of_AMD_graphics_processing_units#Northern_Islands_.28HD_...