CPU Performance Bottlenecks Limit Parallel Processing Speedups

Semiconductor Engineering's Bryon Moyer analyses how hardware optimizations and well-thought-out software architectures can help CPU performance bottlenecks that are currently limiting potential parallel processing speedups.

Multi-core processors theoretically can run many threads of code in parallel, but some categories of operation currently bog down attempts to raise overall performance by parallelizing computing.

Is it time to have accelerators for running highly parallel code? Standard processors have many CPUs, so it follows that cache coherency and synchronization can involve thousands of cycles of low-level system code to keep all cores coordinated. CPUs also have a limited ability to exploit instruction-level parallelism based on CPU width and data dependencies.

These CPU performance bottlenecks are real, pervasive, and not easily resolved. Although software developers have a huge role in creating parallelizable algorithms, there may be room for hardware specifically suited to executing parallel code.