Performance evaluation of many-core systems: Case study with TILEPro64

This study evaluates the performance of the 64-core-based TILEPro64, and compares it with Core i7 and Atom by executing three benchmark programs: a synthetic bench, SPEC CINT2006 and SPLASH-2. TILEPro64 is not advertised for regular applications such as SPLASH-2. However, its internal many-core structure makes it worth investigating the performance characteristic with conventional benchmarks. The synthetic benchmark shows that the stall time because of on-chip network takes up to 85% of total execution time in TILEPro64. The single-core performance with CINT2006 reports that Core i7 and Atom deliver 15.4 x and 3.8 x superior performance to TILEPro64, respectively. The parallel performance with SPLASH-2 reports a similar trend. Comparing the fastest execution times, Core i7 boasts of a 19.2 x faster performance than TILEPro64 and even Atom outperforms TILEPro64 by 2.6 x on average. It came as a surprise that even Atom outperforms TILEPro64 in most of the benchmark programs. The highest number of last-level cache misses is a major culprit for low performance. The forerunner many-core products such as TILEPro64 offer excellent test-beds for polishing, adjusting and reshaping many-core architecture in the right direction.