Replies to post #26251 on Advanced Micro Devices Inc (AMD)

[sgolds]

Tim, I see we both zeroed in on the same issue simultaneously with regards to equivalent translation latencies from PCI-Express in both architectures. :)

[HailMary]

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.

The difference is PCI Express video cards on Intel's platform won't need to go through the CPU bus to get to system memory. On AMD's platform a video card would have to go through HT to get to memory because the memory lives on the other side of the northbridge built into the CPU.

I still argue the small latency penalty for translation from PCI Express to HT is going to make almost no difference in the real world. Video cards don't need low latency access to main memory. They need high bandwidth.

Now that I have stated this in 3 posts, I think I will let it settle out. :)

edit: I see he was claiming something else not related to system memory, with the processor sending a request to a PCI express device. In that case, consider this:

AMD CPU -> On chip connect -> NB -> HT -> HT/PCI Express Tunnel -> PCI Express
INTC CPU -> CPU FSB -> NB -> PCI Express

One less 'hop' for INTC, although AMD's connection between the CPU and NB are on die, and the NB runs at the processor speed may completely negate the extra hop. You could almost treat the AMD CPU+NB as one device:

AMD CPU/NB -> HT -> HT/PCI Express Tunnel -> PCI Express
INTC CPU -> CPU FSB -> NB -> PCI Express