We aren't talking about middlemen like CUDA or OpenCL. We are talking about Real Performance VS. Max Potential Performance. This is a level of testing to see how efficient a GPU architecture is, or even how efficient your software implementation is. Better designed code can lead to better real performance. If your GPU is performing worse in OpenCL than in CUDA, there is a software problem in your coding, or there is a problem with the OpenCL driver with your graphics card. In this case, the problem lies with NVIDIA offering poor OpenCL support with their drivers, while AMD is in the lead with a more efficient driver.

CUDA and OpenCL are software methods that open the door to achieving this performance, but different projects will perform differently on different architectures, so you cannot get 100% potential in the real world. Better designed code, however, can get closer to the max potential of a graphics card. Math projects are simple, easy to code, and reach near max potential easily, but a project like this is more complicated. As it stands, you can count the FLOPS rating based on how fast you can complete a number of work units, which the current rate on my 5 TFLOP machine is 500GFLOPS worth of computations, a 10% level of efficiency. No doubt, this project was hard to code, and in the end, there are still some major dependencies that are being run on the CPU instead of the GPU, which hurts the overall production rate.

If you want to find out how well a project is utilizing your system, divide your average work value in BOINC manager by 200, to know how many GFLOPS are going into that project. For me, that is 100,000, or 500 GFLOPS average.

That was an overwinded discertation ofabsolute nothing! Sorry, as I respect all your posts as a rule. I can also grab stats from equip that would prove all incorrect. STATS bite. Every machine is different, based on so many things. Post as IMO, no prob.

Nvidia

Example about real performance, GTX 465 vs. GTX 650TI, both slightly oced by vendor:
- EVGA superclocked vs. MSI power edition/OC
- June 2010 vs. Oct. 2012
- GF100 vs. GK106
- 352 vs. 768 Shaders
- 1250 vs. 993 Mhz Shader Clock
- 625 vs. 993 Mhz GPU clock
- 810 vs. 1350 Mhz Memory Clock
- 256 vs. 128 Bit Bus width
- 200 vs. 110 W TDP
- 70-80 vs. 40-45 degrees temperature @crunching
- 2300 vs. 1100 rpm GPU cooling fan
- 109 vs. 52 Watts real difference comsumption (Idle > 1xGPU/1xCPU HCC1)
(- 189 vs. 132 Watts comsumption 1GPU-system (BOINC 7xCPU HCC + 1GPU/1xCPU HCC1; undervolting GTX 465@0,900V >165 Watts))
(- missing: no undervolting GTX 650 TI unless valid results)

Both installed in the same system (Slot 1+2 means device 0+1; sharing > PCIe X8, Version 2.0 vs. 3.0 should not show any effect), running simultaneously, same conditions (software, driver, CPU Clock i7-3770K@4,0 Ghz / offset -0.175V). Screen is connected with iGPU (HD 4000), Nvidia 309.97, Win7/64, Z77X-Board/PCIe 3.0 . No modifications with app_info, but cc_config: <use_all_gpus>1</use_all_gpus> for make them running, 260-265 watts peak system consumption with both cards @BOINC.

With HCC1 the new GTX 650 TI takes nearly the double time for the same job, Screenshot with 3xGPU-Z, GTX 465 vs. GTX 650TI calculating 1x workunit HCC1/7.05:
- 15 vs. 8 finished results (Screenshot selected results )
- 7-8 vs. 14-15 minutes each result

In the end, GTX 465 is huge faster, GTX 650 TI runs without any noise (tested it before), that was expected, phantastic cooling made by MSI. Performance of the new GTX is unexpected low, explanation ? The GTX 650 Ti should work better in GPUgrid or other DC jobs.

P.S. Link for comparison GTX 465 vs. GTX 650TI according CLBenchmark

Even though OpenCl can be used on both GPUs, here's the thing:

If you want a complex task to be worked on, it tends to be better to use CUDA. Especially since when you code with CUDA, you usually get the help of NVIDIA to help along the way where necessary.

If you have a simple task, use OpenCl. Just dont expect much help from AMD/ATI or anyone besides a OpenCl programmer and the like.