kernel

Fix scheduler bugs. One more fix pending. Poop.

Apple

Damn I love this Apple. Both the PowerBook and the underlying operating system are a pleasure.

So Apple finally unveiled the G5. 64-bit. Sick HyperTransport interconnect. Half clock speed system bus. Currently up to 2GHz. Runs dual. Amazing SPECfp and - more importantly - great SPECint. Way to go, Apple.

SPECint vs. SPECfp

Speaking of SPEC scores... I see a lot of people on slashdot and elsewhere exclaiming how amazing the SPECfp numbers are for their favorite architecture. Wild claims like SPECfp is all that matters. What crack are these people on? I do not deny for one moment that SPECfp is important, especially for scientific computationally-intensive applications. For example, if I were designing a cluster for physics research, I would pay a lot of attention to SPECfp numbers. I would look at chips like Alpha and POWER and even IA-64 with good SPECfp numbers. But the vast majority of users - and I am talking nearly all desktop, server, embedded, and even many computational powerhouses - are much more dependent on SPECint numbers.

Look at x86. It has awful floating point performance but (with its high clock speeds) excellent integer performance. And, guess what? It performs best at the stuff people care about. So I think some of us need a reality check. For most uses, SPECint is what matters.