Q+A CENTER

To reduce area and power requirements, single-precision floating-point computation is emphasized, which supports the primary PS3 requirements for media and three-dimensional graphics. Similarly, Cell only supports a subset of IEEE floating point arithmetic. While there has been some talk of Cell-based supercomputers, that greatly depends upon the supercomputing application since much traditional supercomputing requires double-precision floating-point arithmetic.

“The Cell chip is general purpose compared to many of the alternative approaches to solving the problems, but it is also special-purpose,” said Craig Lund, CTO at Mercury Computer Systems (Chelmsford, Mass.). Mercury has made a significant, early commitment to Cell and is in the process of launching a variety of hardware solutions. “What Cell is really very good at, is very fast, single-precision floating-point calculations. That capability has a very big footprint on the market, but there are algorithms that do not care about that.”

One noteworthy area of he Cell architecture is how it handles memory. The eight SPUs have individually dedicated local memory. The Synergistic Memory Flow Controller serves as a high-performance data transfer engine, which performs bulk data transfers between the SPU local storage and Cell system memory. By offloading data transfer onto dedicated data transfer engines, the architecture allows data processing and data transfer to proceed in parallel and supports advanced programming methods such as software pipelining and double buffering.

“In the case of Cell, we made a fairly radical change in having a separate DMA engine that actually brings data from shared memory into local memory, where the processor operates on it,” said Hofstee. “It seems like a very simple change in a number of ways, but it has a very dramatic impact on the performance of a number of areas, including imaging. It also applies to things that are more in the high-performance computing space like fast Fourier transforms, and we believe it will likely apply to computations with sparse matrices. And, there are streaming applications. In the very beginning we actually referred to Cell as a streaming processor, and certainly it does extremely well with streaming work loads.”

CELL STATISTICS:

• Observed clock speed: > 4 GHz
• Peak performance (single precision): > 256 GFlops
• Peak performance (double precision): >26 GFlops
• Local storage size per SPU: 256KB
• Area: 221 mm²
• Technology 90nm SOI
• Total number of transistors: 234M

Given the anticipated PS3 volumes, Cell will be a fairly inexpensive system component. Cell is also scalable. The number of attached SPUs can be varied to achieve different power/performance and price/performance points. Multiple Cell processors can also form a symmetric multiprocessor system. In addition, Cell is OS neutral and supports many different operating systems including Linux and embedded.