This happens on at least PIII (Katmai) SMP machines with kernel-2.2.14-5.0 (the pre 6.2 one from rawhide, but the 6.2 version is not different with respect to that (btw: why didn't you increment the release number between rawhide's 2.2.14-5.0 and zoot's 2.2.14-5.0?): Under seldom circumstances floating point calculations yield NaN although they haven't before in the same run. For example the same calculation yields the correct result for several million times when it suddenly "Nans". Maybe sometimes floating point calculations even yield a wrong but non-Nan result, we haven't checked that extensively. We have observed this behaviour on several machines, so it shouldn't be a hardware problem. The processors are 500 MHz Katmai with CPU id 0672 and microcode patches 000B and 0010. It seems to be a problem with the fxsave and fxrestore fast extended FPU save and restore feature used in kernel-2.2.14-5.0's linux-2.2.12-PIII.patch: if one rebuilds the kernel with CONFIG_X86_FX unconfigured the problems disappear. Cheers Cord
The errata kernel doesn't include the PIII patches and as such solves this problem. It's also beleived to be fixed internally in the latest version of the PIII patches.