Last week StarCore unveiled the first details of its next-generation DSP core family, the SC3000. The most notable new feature is the family’s high clock rate: According to StarCore, SC3000 cores will achieve a clock rate of 1 GHz in a high-performance 90 nm process. In comparison, StarCore claims that SC2000 cores can achieve clock rates of about 600 MHz under the same conditions. (The clock speeds quoted in this article assume nominal silicon under worst-case voltage and temperature variations. In contrast, chip designs usually assume slow silicon under worst-case voltage and temperature variations. StarCore has not announced SC3000 clock speeds for slow silicon, but these speeds are likely to be 20-30% lower than speeds for nominal silicon.)
The higher clock speed is enabled by use of a longer pipeline. The SC3000 will use a twelve-stage pipeline—twice the length of the pipeline in its predecessor, the SC2000. The SC3000 pipeline will also be longer than the pipelines in most competing processors.
Although the long pipeline enables a higher clock rate, it also presents new challenges. Processors with long pipelines typically have long instruction latencies. Among other problems, these long instruction latencies can reduce processor efficiency and complicate code optimization. Fortunately, StarCore has taken steps to address these challenges. Most importantly, StarCore has maintained single-cycle latencies for load-execute and execute-store sequences.
StarCore has also boosted performance with a variety of new instructions. The most notable of these instructions are new instructions that accelerate Viterbi decoding, a key algorithm used in communications applications. According to StarCore, these new instructions reduce the cycles needed for Viterbi decoding by a factor of 2.5. BDTI has not evaluated this claim, but it has implemented the BDTI Viterbi Decoder benchmark on the SC1400. BDTI found that the SC1400 was very efficient on this benchmark—only a few other processors can complete the benchmark in fewer cycles. Improving this result by a factor of 2.5 would be a remarkable accomplishment.
The SC3000 cores will also add a memory management unit (MMU) that will enable the cores to run full-featured OSs such as Linux. This feature will give the SC3000 a significant advantage because few other DSPs include an MMU. The SC3000 will be particularly attractive for some applications that currently require both a DSP and a general-purpose processor. In such applications, it will often be possible to replace both cores with a single SC3000.
Although StarCore has not released the full details of the SC3000 cores, it is clear that these cores will greatly increase the capabilities of StarCore’s lineup. StarCore plans to announce specific SC3000 cores in mid-2006. BDTI looks forward to performing a more detailed analysis at that time.