AMD Sees Future in Accelerated Computing

 
 
By Scott Ferguson  |  Posted 2008-02-06
 
 
 

AMD Sees Future in Accelerated Computing


SAN FRANCISCO -When it comes to the future of multicore processing, Advanced Micro Devices is looking far beyond adding more and more transistors to each new generation of chips.

With software development lagging behind the advances in multicore processing, Chuck Moore, an AMD Senior Fellow, said the company has begun exploring new ways that other pieces of hardware-or accelerators-can be combined with traditional CPUs to increase the performance of software applications, while allowing these applications to take full advantage of the new processing technology.

Moore presented his views on developing multicore technology at a panel discussion Feb. 6 at the 2008 International Solid State Circuits Conference here, which also included comments from other top engineers from such chip makers as Sun Microsystems and Intel.

When multicore technology first came into the market, operating systems and other third-party applications were able to take advantage of the extra computing power. However, in an interview with eWEEK before his presentation, Moore said that as the number of cores in each chip grows, users now want their applications to run as fast as the new generation of multicore processors will allow.

The problem, which has been building for years, is that the software being written now does not take advantage of parallel processing. AMD is not the only chip maker to see this problem developing. Intel, which continues to dominate the x86 chip market, is looking to create tools that will help software vendors and developers support application development for multicore systems.

For Moore , the problem with current software development is clear cut. Hardware, specifically PCs, that are built with multicore chip platforms, but that use the type of software still being written today, will show a decrease in performance.

While the number of transistors-the building blocks of microprocessors-continues to increase year after year, Moore said that does not translate into better performance.

"During the past 30 years, we have translated that the more transistors we place on the chip increases performance," said Moore, referring to Moore's Law, which is named after Intel founder Gordon Moore and states the number of transistors on a chip can double every 18 to 24 months. "That has been the main driver of adding value into the chips mainly because we have been able to translate those transistors into increased performance. What is confused with Moore 's Law is that I'm going to double my performance every 18 months."

The solution, he said, is start developing smaller pieces of hardware or subsystems to help increase performance and allow the client, whether a desktop or a notebook, to take advantage of x86 processors that are multicore.

AMDs acquisition of ATI


The most obvious example of this is adding a GPU (graphics processing unit) to the processor package to help better render video and other multimedia applications. AMD has already announced that it plans to combine CPUs and GPUs on the same piece of silicon in 2009 under a program called Accelerated Processing Units, previously referred to as "Fusion."

AMD has another initiative called "Torrenza," which looks to spur the development of co-processors for systems that use AMD 's Opteron processors.

One reason why AMD acquired ATI in 2006 was to take advantage of the company's graphics portfolio as it moved toward this type of chip development. In this case, the GPU allows the software's instructional threads to run in parallel, breaking the information down into smaller pieces to process them simultaneously, which provides for high throughput and better performance for various applications without relying on increasing the clock speed to increase performance. The result of all this is what Moore calls a heterogeneous microprocessor that has a combination of GPUs and CPUs working together, which should increase performance while reducing power consumption. It also allows applications to take better advantage of the multicore platform.

"It is simply that hardware with a specific purpose is much more efficient," Moore said. "You wouldn't want to decode video on a CPU. You want to decode that video on a dedicated piece of hardware that is off to the side. At the same time, it achieves the same performance at one-twentieth the power."

AMD is moving toward offering these types of subsystems for its chips-GPUs are just one example-that are mostly geared for desktops. Moore said most data center servers run applications that can currently take advantage of multicore technology, such as Web services and financial processing. By 2009, AMD plans to offer its first platform that will support chips that have up to eight processing cores.

That's not to say that AMD is only focusing on chips for PCs. In November, the company introduced its FireStream 9170 GPU for the HPC (high-performance computer) market. In this case, it's focusing on the HPC market, where Nvidia, also a vendor of GPU technology, has taken an interest.

Rocket Fuel