Results are available for kernbench, SDET, and SPECjbb on a 32-way NUMA box. CONFIG_SCHED_SMT was enabled, as well as CONFIG_NUMA. Data samples were taken every 5 seconds via "cat /proc/schedstat". On some loads this could create its own biased view (since it must create and execute processes to collect the data), but none of these benchmarks are believed to be that sensitive. Pages and graphs (except for this one) were created by a set of tools soon to be made available.
Kernbench results are a little spotty and hard to extrapolate from, since a kernel compile on a 32-way finishes quite quickly now. There's all of ten data points to draw the graph.
SPECjbb results offer a little richer look into the system behavior, since it took nearly 20 minutes to run.
SDET results took even a little longer to run.
Textual, rather than graphic, results are also available for kernbench, specjbb, and sdet.