<< Is the latency of Rambus high? >>
That's a difficult question to answer, a single RDRAM chip has far far lower latency than SDRAM.
But the thing is with SDRAM it's parallel so latency is fixed, doesn't matter how many chips you add. Rambus is the other way around, it's Serial, so the more chips you add the higher the latency gets. Also there are sleep states and what not for Rambus, lots of factors in it's latency. But I don't think that it matters as much...
i820 and i840 are BAD chipsets, Rambus is not good for P3s. The p6 core has a 64bit front side bus at 66-133Mhz which matches SDRAM. The p6 is more or less designed around SDRAM. The P4 core is designed to make use of Rambus' super bandwidth. It has big fast caches to help mask some of the latency, and it has a monster FSB equivilant to a 400Mhz P3 bus. The P4 dominates at Quake3, many people laugh that off as one game, without examining the why. Q3 loves Bandwidth, both Video card and CPU bandwidth. It's not SSE2, and it's damn sure not "Double Pumped" ALUs that make the P4 good at Q3.
A lower latency/lower bandwidth RAM type (SDRAM) would mangle the P4's performance, much like a higher latency/high bandwidth RAM type (RDRAM) mangles the P3's performance. They aren't designed to each other.
So the short answer is yes, RDRAM does in normal circumstances have higher Latency than SDRAM, but I don't think it matters.
Just so you know my stance: I'm not a Rambus-zealot, I think Rambus Inc (the bunch of lawyers) can go f*ck a monkey, however RDRAM (the memory type) has good potential for the future.