Haddock, Re: Amazing what you can do when the entire benchmark fits in on-die cache!
If that were the case, how does Power4 not do any better, given that it has 128MB of cache at its disposal?
Or to argue another way, consider the scalability of Madison over McKinley. 2x the cache and 50% more frequency yields about 46% in SPECfp. If the benchmark suddenly fit inside the cache where it did not before, then you would expect a much higher score, since cache latency is an order of magnitude better than memory latency. On the other hand, if the benchmark fit inside the cache in McKinley as well as Madison, you would expect a lower score, since SPECfp does not scale very well with frequency alone. The 46% improvement tells me that there is still a significant amount of the benchmark that resides outside the cache, but a higher hit rate allows better scalability than just the raw frequency boost.
Market Data 
Markets 
AI
