Depends on whether they’re talking about single precision or double precision operations...I assume the latter.
Right now a high end quad-core CPU is good for about 70 gigaflops (.07 teraflops) of double precision operations.
A high end GPU, which performs single precision calculations, is good for 900+ gigaflops (almost a teraflop). Way faster, but only good for very simple, repetitive calculations.
- Floating Point. The CELL processor used in the PS3 has good single precision (32-bit) floating point performance, but is not very good at the double precision FP used in most scientific/engineering codes. This is why the CELL (The PowerXCELL 8i) processor used in the RoadRunner machine is uses a different design than the PS3 CELL. It was redesigned to perform double precision floating point more quickly (they also changed the memory system to use DDR memory instead of XDR and increase capacity).