macstl offers a new implementation of std::valarray optimized for SIMD on PowerPCs and Pentiums. It runs 3.6x to 16.2x faster than hand-coded scalar loops. A portable SIMD vec class lets you work with low-level SIMD instructions equivalently cross-platform.
The Mach copy-on-write vector is a partial specialization of std::vector for the Mach allocator. It neatly uses the OS vm_copy facility to "get around" thread safety issues when doing copy-on-write and reference counting. As a result, std::vector can be copied, passed back by value and enlarged without the same overheads.