The design of experimental systems usually employ some form of simulation. The inputs to the simulation are typically standard benchmarks. This paper presents a method for finding a cost-effective design for each benchmark. This method is called benchmark characterization. Benchmark characterization is possible in part due to the advent of high-performance architecture-independent compiler technology. To demonstrate the method, seven benchmarks are characterized. Five of the benchmarks are from the SPEC benchmark set (gcc, espresso, spice, li and matrix300) and two are popular synthetic benchmarks (dhrystone and whetstone). Benchmark characteristics are reported for the processor, memory system, and operating system