H3nrik V! wrote on Today, 05:39:

[...] […]
Show full quote

[...]

Yes, that's a funny note - but you wrote it yourself: "I guess they did the whole "move the northbridge on-die" in steps."

If I look at it, it seems like the 1156 launched close to a year later than 1366 did. I guess the 1366 was the first "flagship" on-die memory controller chip, where they only put the memory controller on-board, rather than everything all at once. The thing is that the on-board PCIe on 1156 was on the expense of "only" dual channel memory. (not that it really made any practical difference - but data sheets are data sheets 🤣 )

I guess the fact that the QPI link was scalable (4.8GT/s to 6.4 GT/s) made it possible to also scale the number of PCIe channels from the northbridge side, whereas the 1156 had one x16 interface. That was probably not enough for the HEDT platform in Intel's mind back then. Also, if I understand correctly, the QPI is also used on Xeons for inter-socket communications?

Not so strange at all if you look at the context.

LGA1366 was originally designed for the Xeon server platform, but - as several times before and since - Intel decided to introduce it in November 2008 as a high-end desktop platform as well. As with other platforms primarily intended for the server market, features were more geared towards bandwidth than latency. So big wide QPI path to an extermal chipset, where you could basically deploy as many PCIe lanes as you want. Sure, latency would be higher than if using a dedicated PCIe controller in the CPU, but for the workloads it was designed for, that was hardly an issue.

LGA1156 was a dedicated desktop platform with some nods and winks to laptop development. It not only moved the memory controller to the CPU like in LGA1366, but also included the northbridge and so the PCIe controller. The remaining chipset component - the PCH - was essentially a southbridge in all but name, and the DMI link between CPU and PCH was basically just PCIe as well.

The end result of the LGA1156 and later integration was a more modest bandwidth than LGA1366 offered and no sensible way to implement SMP, but with significantly reduced latency, so ideal for desktop use. These very different characteristics led to a divergence in benchmarks: things that thrived on raw CPU power and memory and PCIe bandwidth did a lot better on LGA1366 platforms, but things more sensitive to latency performed better on LGA1156. That meant that when it came to gaming, with a single GPU the LGA1156 scored better, but add a second GPU and LGA1366 came on top.

You saw similar things with later server->desktop platforms like the LGA2011.