I would expect this to scale well with at least the first few additional cores: the total memory bandwidth used in this test is probably 10 GB/s or less, wherea

I would expect this to scale well with at least the first few additional cores: the total memory bandwidth used in this test is probably 10 GB/s or less, whereas system memory bandwidth depends on the number of memory channels and memory speed but ranges between 30 and more than 100 GB/s for most configurations, so there is a fair amount of unused capacity at the memory subsystem. With enough additional cores though you'll hit some kind of shared limit, however. How many depends strongly on the details on your system, and this varies by nearly a factor of 10 between some server and client systems.