Santa Clara, Calif. -- Delivering throughput of as much as 31 gigamultiply-accumulates per second (GMACs), a crop of processors unveiled at last week's Multicore Expo could make short work of complex compute-intensive algorithms in wireless basestations, medical imaging, network routing, video transcoding and multimedia. They also show progress in exorcising the demons that have dogged the category thus far.
Conference presentations here detailed developments in multicore and multithreaded architectures and highly parallel-processing arrays, but success for such advanced processors is not a given. The goblins range from a dearth of programming and development tools to the specter of industry consolidation.
Algorithm complexity has been on the rise in communications and multimedia applications as designers work to make signals more robust, enable the recovery of data lost to channel noise or to failed bits in the storage medium, and allow compression and playback of image and audio streams without unduly compromising quality or crippling real-time capability. The good news is that many of the computations can be done in parallel, letting systems tap multiple processors to handle the workload. The bad news: All those processors must be programmed and synchronized to deliver the processed data.
Many of the reconfigurable multiprocessor chips that have been brought to market thus far have fallen by the wayside, victims of architectural shortcomings, cumbersome programming tools or funding economics. "The latest batch of vendors to develop architectures may be in for some consolidation," Will Strauss, lead analyst at Forward Concepts (Tempe, Ariz.), said in a presentation. "The current crop of 20 or so vendors might shrink down to about five suppliers over the next five years, even with new companies continually emerging."
The challenge, Strauss said, is not just in building higher-performance silicon, but also in being able to program it to execute the required algorithms. Many of the early multiprocessor vendors failed to deliver on that last count. Some tools required too great a learning curve. In some cases, the processor vendors had to develop the application software on behalf of the customer. That did not sit well with many potential customers, who preferred to keep their algorithms--cherished intellectual property--close to the vest.
As software becomes the differentiator for many system companies, development and programming tools are increasingly vital components of any project. That's reflected in the growing number of application development teams on which software developers outnumber their hard- ware counterparts. The latest crop of multicore chips is coming to market with better tools than the ones that supported their predecessors. And many third-party suppliers have crafted chip-independent tools to ease software creation for multiprocessor architectures (see story, page 1).
Simple to sublime
The range of silicon solutions discussed at the conference spanned from simple heterogeneous approaches to highly integrated processor arrays on the bleeding edge of performance.
In the former category was an advanced sound subsystem processor from ARC International that combines a high-throughput, 128-bit-wide single-instruction/multiple-data (SIMD) processing array and an ARC 700 series CPU core. The combination yields overall compute throughput of more than 9.5 giga-operations/s when clocked at 500 MHz.
The processor adds a library of audio codecs to the SIMD engine and a two-channel audio-optimized direct-memory access controller to manage the data streams. "This lets the subsystem deliver a high-speed transcoding capability as well as high-speed audio encoding and multistream high-fidelity audio decoding," said Derek Meyer, senior vice president of sales and marketing at ARC.