accord99, there's no point of arguing with you any more. I clearly see that you do not pursue the formal architectural approach, but just rant along from an enthusiast point of view.

When you start to respond only to my immediate comments above, that's when I know you're not making any point at all. There's simply no way to make you look at the whole picture. It's said that the PC market (or probably most enthusiast markets) is filled with this type of religious thinking.

From my first-order collection of what we've argued, I'd only say a few things:

1. For 4S or below, snooping via cHT has very low latency. The snooping message goes in parallel with memory access. It is the remote access that's high-delay, but if this happens a lot, it's indication that the program is not NUMA-optimized.

2. Lots of snooping traffic will impact multiple cores that do not shared a cache. This is true to UMA and FSB. In fact, this increase contention on the bus, and 2x contention is more severe than 2x data rate in terms of bandwidth efficiency.

3. Memory access speed increases ~2x per 10 years. Pin count is ~4x, and bus speed ~8x, processor speed and memory parallelism ~16x.

4. If you don't like virtual machine, or don't like multiple sockets, don't use them. Nil would I care. Stop pretend that nobody would want them, and stop spread FUD and incorrect implication on NUMA or snooping.