Print 84 comment(s) - last by dwalton.. on Jul 16 at 4:09 PM

Intel says CUDA will be nothing but a footnote in computer history

Intel and NVIDIA compete in many different ways. The most notable place we see competition between the two companies is in chipset manufacturing. Intel and NVIDIA also compete in the integrated graphics market where Intel’s integrated graphics chips lead the market.

NVIDIA started competing with Intel in the data processing arena with the CUDA programming language. Intel’s Pat Gelsinger, co-general manager of Intel’s Digital Enterprise Group, told Custom PC that NVIDIA’s CUDA programming model would be nothing more than an interesting footnote in the annals of computing history.

According to Gelsinger, programmers simply don’t have enough time to learn how to program for new architectures like CUDA. Gelsinger told Custom PC, “The problem that we’ve seen over and over and over again in the computing industry is that there’s a cool new idea, and it promises a 10x or 20x performance improvements, but you’ve just got to go through this little orifice called a new programming model. Those orifices have always been insurmountable as long as the general purpose computing models evolve into the future.”

The Sony Cell architecture illustrates the point according to Gelsinger. The Cell architecture promised huge performance gains compared to normal architectures, but the architecture still isn’t supported widely by developers.

Intel’s Larrabee graphics chip will be entirely based on Intel Architecture x86 cores says Gelsinger. The reason for this is so that developers can program for the graphics processor without having to learn a new language. Larrabee will have full support for APIs like DX and OpenGL.

NVIDIA’s CUDA architecture is what makes it possible to process complex physics calculations on the GPU, enabling PhysX on the GPU rather than CPU.

Comments     Threshold

This article is over a month old, voting and posting comments is disabled

By dickeywang on 7/3/2008 5:28:43 AM , Rating: 2
1) Limited on board memory. You can see the problem simply by dividing those Flops numbers posted by Nvidia with current GPU's on board memory size. The GPU itself can process floating point data really fast, but high performance computing usually also requires very large memory size. GPUs Flops numbers is hundreds times the numbers of CPUs, but when considering RAM per GPU/CPU, they are comparable. So if the memory requirement of your simulation requires 4000 CPU nodes, you will also need 4000 GPU nodes if you use CUDA.

2) The reason GPU is more powerful than CPU when dealing with raw floating point data is because GPU has so many stream processors. However all these stream processor share the same memory, so if you are only doing simulations on one GPU, you need to write a code in a "share-memory" way, however, if you want to do some serious numerical work, you probably need to have hundreds or more GPUs running parallel, which means you have to deal with the local "share-memory" nature of each GPU, but also deal with the parallelization of your code for the entire cluster. Looks to me it would require a "OpenMP+MPI" types of structure of code.

I think if Nvidia wants CUDA to be popular in the high performance computing community, these two problems need to be addressed. If (a large portion of) these two problems can not be dealt with by the compiler itself but requires a lot of work by the programmer, it will be tough for the high performance computing community to accept CUDA.

"People Don't Respect Confidentiality in This Industry" -- Sony Computer Entertainment of America President and CEO Jack Tretton
Related Articles
GeForce 8 To Get Software PhysX Engine
February 15, 2008, 10:33 AM

Copyright 2016 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki