Ah - is this a quad-core CPU? Interestingly there does seem to be a pattern building up here, which is that it's more SKU specific than I previously thought - it now looks like it's limited to the S and U parts, which are paired with the Sunrise Point-LP PCH. CPUs that are using Sunrise Point-H appear to work more as expected.

Thanks, I'll try to replicate your PCI configuration and see if I can trigger any behavioural changes.