Replies to post #26248 on Advanced Micro Devices Inc (AMD)

[Tim Fowler]

The latency to make one HT hop looks to be about 40 ns. Intel's northbridge latency is probably also in that vicinity. But now, here's the rub. HT needs to be converted into PCI Express. There's a translation latency as well as the standard PCI Express latency. In other words, Opteron has more latency when outputting a PCI Express signal than Intel's system.

I don't think Intel's Northbridge talks to the CPU in "PCI Express". Just as the signal on PCI-E needs to be converted to the type of signal used in HT the signal on PCI-E needs to be converted to whatever method or protocol Intel uses to communicate from the northbridge to the CPU.

Tim

[sgolds]

calbiker, I am glad we are in substantial agreement! There is something that I don't understand from your issue:

But now, here's the rub. HT needs to be converted into PCI Express. There's a translation latency as well as the standard PCI Express latency.

1. Isn't there a similar translation latency in Intel's North Bridge to convert from PCI-Express to the processor bus format?

2. So far we are talking about the idealized case where the only thing happening is a PCI-Express access. Now, what happens when you mix in an overlapping memory access from the processor? Seems to me that the P4 will suffer a degradation in latency and throughput because both requests & both data accesses happen on the same processor bus (vs. Opteron, which has a separate memory bus).

By the way - don't know if you noticed my other post replying to the suggestion made today that a future Intel North Bridge may incorporate the PCI-Express fabric. In that case, P4 video would greatly benefit (route to DRAM bypasses the processor), although Opteron will still have a clear edge in memory-to-processor access. In that theoretical matchup the question of who's faster would be highly application dependent.

[Dan3]

Re: HT needs to be converted into PCI Express

Are you sure about that?

PCI Express may be DOA as anything more than a limited use near proprietary bus (microchannel, anyone?).

Test results, to date, show it to be inferior to AGP - and, unless and until it can become high volume, it's going to be more expensive, to boot.

Sounds a lot like rambus, and I'm sure you remember what a dismal failure that one was, even with Intel pushing it just as hard as they possibly could.

http://www.onlynewz.com/modules.php?name=News&file=article&sid=1516