The NAS Parallel Benchmarks (NPB) from NASA is always interesting with their very efficient OpenMP-based multi-threaded tests written in FORTRAN. Running the IS.C test set in a virtual machine costs about 30% of the performance at least for the Intel Core i7, but there was virtually no change in performance between KVM and VirtualBox.
While the NPB IS.C performance was indifferent between KVM and VirtualBox, with NPB LU.A, KVM with the Linux 2.6.35 kernel in Ubuntu 10.10 was 20% faster than the latest stable VirtualBox release.
Lastly, with the NAS Parallel UA.A test configuration, KVM was 40% faster than VirtualBox 3.2.12 on this twelve-thread Intel 64-bit system.
In a majority of the benchmarks used that were mostly CPU oriented, the Kernel-based Virtual Machine was measurably faster than VirtualBox 3.2.12 and VirtualBox 4.0 Beta 2. The only area where VirtualBox decisively pulled in front of KVM was with the disk benchmarks, but it's a situation where VirtualBox by default isn't obeying sync/fsync requests made within the guest, so it's numbers are shown as even being faster than the host operating system, but it comes with potential data integrity issues.
While VirtualBox may be slower than KVM in many areas, VirtualBox does provide 2D and 3D acceleration support to guests, among other features, not currently offered by KVM/QEMU.
As far as the VirtualBox 4 performance goes, in many areas right now its performance with the just-released Oracle VM VirtualBox 4.0 Beta 2 is slower than VirtualBox 3.2.12. However, it is a beta release not intended for use in a production capacity and we will be back with more VirtualBox 4.0 benchmarks upon the program's final release to see whether these regressions are still present or how the performance shakes down.