NVIDIA Not Especially Set on Optimizing CUDA for x86 Chips

It seems that, even despite a certain recent announcement, the NVIDIA CUDA technology is never truly going to run perfectly on x86 chips, developers being, at least for now, limited to just testing the reliability of the CUDA-based software approach.

Some end-users may be aware of the fact that, not too long ago, NVIDIA released a compiler that lets CUDA applications be accelerated by x86 CPUs.

While this definitely expands the compatibility of said applications and may look like it has certain advantages, analysts appear to think that software makers will still have a lot of work to do.

The simple fact is that CUDA applications are especially designed for NVIDIA GPUs, which means that, compiler or not, they will run much more slowly on x86.

There is also the possibility that the new compiler does not support such features as the AVX that MAD's Bulldozer and Intel Sandy Bridge CPUs come with.

This lack of support would be no different from how GPU-based PhysX tools don't support SIMD extensions like SSE2.

The compiler will let developers test CUDA-based software on x86 platform, but the reliability of this approach has not been established yet.

All in all, according to Jon Peddie Research analyst Alex Herrera, the performance gap between GPU-based and x86 CUDA may just be too high.

"CUDA on x86 is going to be slower than an application optimized to run on x86 without CUDA, probably a lot slower,” Herrera stated.

“So a developer running a CUDA application on x86 and then on Fermi is going to see a larger speed-up than he might otherwise have had had he first optimized on a conventional, non-CUDA x86 platform,” he added.

“Bigger speedup numbers serve Nvidia’s purposes of showcasing how much faster GPUs are than CPUs on many floating-point intensive applications," the analyst went on to saying.