Gaidheal wrote:
> "How did you arrive at that 90% value?
>
> LOL I did not.. the people whose job it is to look at things like
> that did :�)
Which would be, who? Just because someone has a web site doesn't mean they are an 'expert' nor, no offense intended, do I know if you interpreted it correctly.
> Why not do the same research I did to find it? Hehe!
Because, for one, I don't know what kind of search you did and, since you 'already did it', I would think you could provide the source.
>
> Conservative estimates have always been "50%" since the days of
> "crossing the 1GHz barrier".
Again, I need a source for those "conservative estimates" to know who is making them, their qualifications to do so, the method they used to determine it, and precisely what it means.
>
> Of course, SOME applications are going to keep the CPU working more,
> but even there, the CPU is not the bottle neck, memory and other I/O
> devices are. A modern fast CPU spends a lot of time waiting for data
> to be retrieved from memory, even with improvements in cache
> technology, prefetching, etc.
There IS a "lot of time" (whatever that means) spent 'waiting' for memory but one of the most significant chunks is memory latency, and no amount of 'bandwidth' will solve that (faster access times would help and longer memory streams would help [as in larger cache sizes, etc]). Which is but one example of what I mean by asking "what does it mean?" when a generic statement like "spends a lot of time waiting" is tossed out. e.g., if the 'time waiting' in the articles you read are due to memory latency then the suggestion to increase FSB vs processor speed isn't necessarily good as latency will remain the dominant factor.
Just off hand, what you seem to be implying by "spends 90% of it's time waiting," and that it is due to the memory pipe bandwidth being maxed out, is contradicted by observable performance numbers. If it were true then increasing processor speed would have virtually no effect on performance as performance would be completely memory pipe bound (I.E. you're only able to execute as many instructions as come down the already speed limited pipe, which can't provide 'more' of them) and you would just be processing the previous 10% portion faster while waiting more. (This is analogous to the jackrabbit start dragster driving down a street where the lights are timed for 30 MPH. He just gets to the next stop light quicker [time spend 10%] and then waits at the next one longer [time spent 90%] than the more sedate driver. And him going faster just means he get's there quicker to wait longer. I.E. The "hurry up and wait" syndrome.)
However, even memory intensive benchmarks (meaning high data throughput as well as instructions stressing the memory pipe) show corresponding performance increases with CPU speed (within the realm of currently available CPU to memory bandwidth ratios) so the memory pipe is apparently not 'maxed out' with the slower processor. Otherwise, anyone who buys more than the slowest, cheapest, processor available on any particular speed FSB is a fool as they get 'nothing', under your theory, for more processor speed.
