Here are the results:
The popular SuperPi benchmark is a good reflection of the performance of a computer and it shows right away that the reduction of the L2 cache to 128KB doesn抰 lead to any catastrophic consequences with K8 processors. The performance degenerates but slightly: the CPU with a 128KB L2 cache is only 2% slower than the model with 1MB of L2 cache memory. On the other hand, SuperPi isn抰 very sensitive to the speed the processor is working with the memory at, so let抯 see what we have in other tests.
PCMark04 doesn抰 reveal any serious differences between K8 processors equipped with different amounts of cache memory. In the overall rating, and in the CPU rating, the processor with the smallest L2 cache is less than 2% worse than the model with the biggest one. The memory subsystem performance subtest finds the difference somewhat more dramatic: the Sempron with 128KB of L2 cache is 4% worse even than its analog with 256KB of cache memory.
3DMark2001 SE yields very curious results. As you see, a perceptible performance gain occurs when the size of the L2 cache grows from 512 to 1024 kilobytes. As a result, while there抯 a one-percent-or-less difference between 128KB and 256KB of cache, the gap between 128KB and 1MB of cache is more than 11%.
Unlike its ancestor, 3DMark05 doesn抰 appreciate big amounts of L2 cache memory much with regards to CPUs of K8 architecture.
Gaming applications seem to be the most sensitive to the size of the L2 cache. The performance gap between processors with 128KB and 1024KB of cache can be as wide as 21% (in Doom 3). In average, the Sempron with a 128KB L2 cache is about 12-13% slower than the Athlon 64 with a 1MB L2 cache in games, if they are clocked at the same frequency.
Note the curious fact that the Sempron with 128KB of cache is faster in Quake 3 and Unreal Tournament 2004 than the analogous Sempron with 256KB L2 cache! This paradox is easily explained: the sample of the processor with 128KB of cache memory was based on the 90nm stepping D0 core, while the rest of the participating CPUs were based on 130nm CG stepping cores. The D0 core features an improved integrated memory controller which sometimes provides certain performance gains. This gain compensates for the reduced cache with interest in Quake 3 and Unreal Tournament 2004!