by Anton Shilov
07/22/2010 | 01:24 PM
David Kirk, an Nvidia Corp. fellow and the former chief scientist of the graphics company, admitted that heterogeneous computing architecture is the most efficient since it allows to process all types of data in the best way. But Mr. Kirk is not sure that Intel Corp., which is developing its multi-core code-named Knights Corner accelerator for high-performance computing will succeed in creating a viable heterogeneous multi-core platform.
At present many companies working in the fields of oil and gas exploration, seismic processing financial services and other are employing graphics processing units (GPUs) and/or special compute accelerators or their base (such as AMD FireStream and Nvidia Tesla) for high-performance computing (HPC) instead of traditional central processing units (CPUs). Intel Corp., the world's largest maker of microprocessors, failed to deliver its own graphics chip code-named Larrabee and is currently working on the code-named Knights Corner chip that will combine many-core architecture with x86 compatibility. Still, in order to run modern operating systems traditional CPUs will be required and Mr. Kirk does not expect them to disappear.
"We find that most problems, if not all, are a mix of serial control tasks and parallel data and computation tasks. This is why we believe in heterogeneous parallel computing - both [parallel and serial] are needed. CPUs are commodity technology and there are multiple CPU vendors that we work with. In my opinion, Intel has no particular advantage in developing a hybrid system - in fact, they have had little success historically in designing either parallel machines or programming environments," said David Kirk on Thursday during a public interview.
Intel's HPC platforms featuring Knights Corner will consist of separate CPUs and many-core HPC accelerators that will be plugged into PCI Express sockets. Many HPC specialists believe that PCIe bus is a bottleneck for such accelerators because of low bandwidth and one of the things that could solve the problem is creating a chip that combines x86 cores with massively-parallel graphics cores; something that Advanced Micro Devices is doing with its Fusion project. But Mr. Kirk claims that PCI Express' bandwidth is not necessarily a bottleneck.
"Contrary to popular belief, the PCIe bandwidth is not often the bottleneck in most applications. The PCIE bandwidth is faster than many other data paths in the system, including the disk, the network, and in many systems, the system memory bus or front-side bus. That being said, there are certainly technical improvements we can make going forward [to solve PCIe bandwidth potential problems]. You'll have to wait and see," said David Kirk.
Just like other specialists in the field of HPC and parallel computing, the Nvidia fellow does not believe in the future of Cell processor designed by IBM, Sony and Toshiba. Mr. Kirk, who currently focuses on CUDA and GPU computing education and research, claims that Cell has played some role in defining the current realities of the market, but which does not represent a threat to modern GPUs in appropriate spaces.
"The Cell processor was a great innovation for its time. Many of the ideas in Cell, including heterogeneous processing and local memory, are part of a modern GPU compute architecture. Cell was a 'point product' though, which means that it did not continue to evolve. We continue to evolve and improve our GPU architectures every 6 months or so. This makes Cell no longer competitive," said the fellow of Nvidia.