A multi-core, multi-threaded gaming engine described as "a supercomputer on a chip", Cell is expected to be chosen as the chip for the PlayStation 3 and will give ten times the capability of current PC processors, the companies said.
They chose the International Solid State Circuits Conference in San Francisco to disclose the first technical details about the four-year project. The prototype chip comprises one 64-bit Power PC processor core and eight separate processing cores that the companies call "synergistic processing elements", or SPEs. The cores can support multiple operating systems and programming models through virtualisation, said Jim Kahle, director of technology at the Design Center for Cell Technology, and an IBM fellow.
Chip companies have turned to multi-core designs in recent years as the performance of single-core processors has levelled off. A single-core processor can be set up to process multiple instruction threads at the same time, but must ultimately run faster and faster to improve its performance, which generates heat. Multi-core processors can execute instructions in parallel, which means multiple separate instruction threads can be processed at the same time. By moving to multiple-core designs, chip designers can extract more performance from their products while reducing power consumption and heat dissipation.
The Cell designers have figured out how to push both frequency and parallelism, Kahle said. Cell is capable of running at more than 4.5GHz, but the companies have not determined how fast the final product will run. The chip will work with air-cooled designs similar to those used by the PlayStation 2, he said.
Cell will probably consume around 30 watts of power, similar to the Emotion Engine processor in the PlayStation 2 console, said Peter Glaskowsky, a technical analyst with The Envisioneering Group in Seaford, New York. This is also similar to the power consumption of Intel's Pentium M processor.
The dual-threaded PowerPC core functions as a control processor for the other eight single-threaded SPEs, which do the majority of the heavy lifting, Kahle said. The SPEs are designed for floating-point calculations, which are very important in graphics processing and supercomputer applications such as seismic modeling.
The SPEs have 256KB of cache memory on each core, while the PowerPC core uses 32KB of Level 1 cache and 512KB of Level 2 cache, said Masakazu Suzuoki, vice president of microprocessor development at Sony. Cache memory is used to store frequently accessed data close to the processor to avoid the delay associated with retrieving that data from the main memory.
The cores are held together by a shared bus that moves data into and out of the SPEs, Suzuoki said. Data is moved into and out of the chip with memory and I/O controllers that are integrated directly onto the processor, rather than implemented in a chipset.
Rambus designed both interfaces used in the Cell processor, said Rich Warmke, director of marketing for Rambus, in an interview prior to the Cell announcement. The memory interface uses Rambus' XDR (extreme data rate) standard, and the I/O interface uses the company's FlexIO technology, he said.
With Cell capable of handling ten instruction threads at a given time, the chip needs extremely fast memory and I/O buses to ensure the processing engine is continuously fueled with data, Warmke said. The XDR bus runs at 3.2GHz, while the FlexIO bus runs at 6.4GHz, he said.
Cell is primarily designed for digital-home applications, but the chip can be used with a wide variety of programming models and could conceivably wind up in any number of systems, Kahle said. The companies are working with open-source compiler developers to create software development tools for programmers, he said.
The prototype wafer and chip shown Monday were built on a 90-nanometer process technology at IBM's manufacturing facility in East Fishkill, New York. Sony will also make some of the chips at its Nagasaki, Japan, fab this year, the companies said.
As compared to conventional processors, Cell is fairly large, with a die size of 221 square millimeters. The latest version of the Pentium M is only about 84 square millimeters.
A larger processor is generally more expensive to manufacture, given the fact that fewer chips can be cut from a silicon wafer. But the Emotion Engine processor in the PlayStation 2 was also quite large, and it's widely known that Sony lost money on the initial shipments of the console, Glaskowsky said.
The processor shown Monday was only a prototype, and it's likely that the high-volume shipments of the processor will come when the three companies are ready to make chips using a 65-nanometer processing technology, Glaskowsky said. That technology will allow the companies to shrink the chip and reduce their manufacturing costs, he said.