2.66 vs. 2.93 : what's the real difference?

[Quu]

If you actually knew more about Tesla and CUDA, you'd know that virtually all Tesla products were Nvidia 8800GTX cards in parallel without video outputs, though they're using the 280GTX now as well. The distinction is pointless as both are consumer graphics cards.

In fact, the technology behind Tesla, the CUDA language, was accessible on non-modified Nvidia cards as well. However, it also is the primary deterrent to usage as it's a proprietary solution that requires specialized coding.

What OpenCL offers is an industry standard that is compatible across OSes and vendors, providing similar benefits as the OpenAL and OpenGL standards. As was an issue with Altivec and CUDA, the benefits of OpenCL will be highly dependent on the parallelization potential of the task. However, many processing-intensive tasks tend also to be suited for parallelization.

I gotta agree with Sauron on all his points.

OpenCL is the answer to a problem we have at the moment. NVIDIA have CUDA, ATi has 'Stream' both are C variant languages for use on computing general purpose tasks on a GPU (Non-Graphics Work). By having OpenCL we can replace these proprietary standards (Which both companies could decide to alter or not support in future card updates) with a single unified and open approach. Microsoft in DX11 are planning their own rival to CUDA, Stream and OpenCL so it is better we get OpenCL in OS X and finalised if we want it to be the standard of choice if Microsoft have their way we're all screwed.

And GPU's are a lot better suited for Floating Point. For example a 4870 X2 GPU (Single Slot affordable card) is around 2000Gflops of performance. Compared to current 3.2GHz Quad Cores from Intel that are around 125 to 170Gflops And I'm not kidding you can look it up yourself. Now not everything relies on Floating Point performance and I know that, but there are some tasks that gain significantly from it. You quote that simply moving to OpenCL will not give a 100% performance boost and for most applications your right but for Video transcoding, rendering and similar tasks it can provide 100%, 200%, 300% or 400%+ more performance over a current generation Intel processor and that is the facts.

 

[thermodynamic]

30 seconds is a huge difference in the computer world. Even 1-2 seconds quicker is a respectable gain. When the CPU renders or is in use every little second adds up throughout the day.

But like another poster said ifportability isn't a factor at all, buy the mac pro desktops as they are miles faster for CPU rendering/storage solutions and especially for rendering.

http://www.macworld.com/article/139919/2009/04/cto_macpro.html

That shows the 2.66 and 2.93 single core units. After reading the conditions of the tests (note which units were tested with 3GB and which with 6GB and those with the ATi vs those with the nVidia), going "one thousand one, one thousand two" still isn't major. One benchmark I recall reading, the video compression as it had the more lucrative values, had the 2.93 shaving 30 seconds off from processing a 6 minute clip... for a 24 minute clip, that's 2 minutes in net savings and I didn't mention the compressor took 9:38 for just those 6 minutes. 8:59 with the faster processor. Multiplied by 4, roughly 38 minutes vs 36 minutes. For video compression. Of a 24 minute clip. x4 = 152 vs 144 minutes or 2 hrs 32 mins vs 2 hours 24 mins for a 96 minute clip. 8 minutes saved is still doggie biscuits compared to the net time involved for solely rendering, which is for all intents and purposes, still two and a half hours.

And all of that occurs during rendering time only. When setting up the material TO be processed, we won't know any difference there. It does add up, but only if it were to save 60 minutes (3600 seconds) per day would I deem several hundred dollars for such an upgrade to be worth it. (Man, it'd be nice if Apple let the 3.33GHz chip be authorized for use... not much than twice more going from a 2.66 to a 2.93 in the Mac Pro single quad and, I'm guessing, would meet the hour-saved criteria.)

 

[cluthz]

You know this is a thread that is almost a year old?