Tesla news: CUDA vs OpenCL

Nvidia's supercomputer architecture may be cheap and awesome, but it's kinda proprietary and falsely green.

Atomic was in on an Nvidia conference call on Friday that went through a presentation about Tesla and GPGPU technology.

Honestly it wasn’t very interesting, but we did get to ask a nice question about compiling to different languages and the benefit of CUDA over OpenCL.

Tesla is Nvidia’s series of video cards that are designed for high power parallel computing. The hardware comes in card, external box and rack-mount server form, and although they’re based on the Quadro graphics card design, they don’t have video outputs.

Nvidia was trumpeting the architecture’s benefits over x86 and the raw performance increases you see in certain applications that have been coded to take advantage of parallel processing on its Tesla hardware. It’s all very impressive, with anywhere between 18 to 149 fold increases in performance over x86 computation depending on the application and CUDA being taught in universities and industry support and low prices compared to x86 supercomputing clusters.

Which is fine, but bear in mind that nobody really likes X86 because it’s the compounded result of 30 years of hacks that that try to pimp out the architecture, which means it’s bulky and inefficient. Seriously, you still have to go through 16-bit mode to initialise an x86 processor – all in the name of backward compatibility. Yet since it’s in everything people will keep begrudgingly compiling for and using it to do computation. It’s at the crossroads of this and parallelisation where Nvidia is trying to step in and offer an alternative for certain applications.

This is why we’re not buying the side order of “green computing” that Nvidia is using to promote Tesla. When some applications are parallelised and run on the new Tesla architecture, they do so faster and more effectively than they would on the X86 architecture. Promoting green computing in this case is like comparing a Model T Ford to a Bugatti Veyron and concentrating only on emissions output. In the scheme of things, it’s simply unimportant.

After some impressive graphs and performance benchmarks were shown, we asked why programmers would bother coding in CUDA – a proprietary coding language for NVIDIA GPUs – when OpenCL can essentially do the same thing as CUDA and produce code that isn’t locked down to Nvidia hardware which as we all know is outpaced by ATI’s gear every few years.

Nvidia’s response was that CUDA is better able to take advantage of the low level features in its hardware than OpenCL. Common standards make features harder to program for, and CUDA lets you get more performance out of an Nvidia GPU than OpenCL. Regardless, Nvidia’s hardware supports both OpenCL and CUDA, and Nvidia believes there is only a minor implementation difference between the two.

When asked if it planned to make some sort of CUDA to OpenCL cross compiler, Nvidia responded that it would have to think about it, that it wasn’t in its roadmap and that it hadn’t committed to it yet.

If you want to learn more about supercomputers, GPGPU computation and CUDA, you’ll be happy to know that we have written extensively about both before.