Posts by Bob Guy

21) Message boards : Number crunching : I think we should restrict work units (Message 13865) Posted 3 Jun 2006 by Bob Guy Post: for what it is worth, I have error detecting and correcting memory on my machine. I think the one-bit errors do not originate in the memory, the errors originate in the FPU/SSE. It is a known fault of the AMD cpus that the AMD FPU processes numbers differently (possibly less accurately) than the Intel FPU (this is usually overcome by proper program code). It is also a fact that overclocking can cause the FPU to be less accurate (the one-bit errors) for both AMD and Intel. One interesting and not well known feature of the FPU is that inside the FPU numbers are not represented as decimals as you might think. So, of course you say: they're binary! This is not true - the numbers are IEEE format for hardware design reasons. There are decimal numbers that can not be represented exactly in IEEE format. The numbers are 'close enough' for most purposes and special code is usually implemented to minimize error - the usual process is by extending precision and using careful rounding. At any rate, the errors introduced by IEEE format can be exagerrated by one-bit errors at or near the limits of precision.

LHC@home