Originally Posted by rluker5
Hopefully the rebirth of the northbridge isn't too costly in terms of performance. It may even help, idk maybe they've figured out a way to lose some performance bottleneck.
The 3 biggest issues with the classic northbridge design was bandwidth, the fact it wasn't full duplex (could only send or receive, but not both at the same time) and that all external communication had to pass through the socket-northbridge link before getting anywhere else. There was also the issue of latency.
The I/O die solves the bulk of these issues:
1. Infinity Fabric, being a significantly newer technology, is inherently faster than any of the old FSB/QPI/HT links ever could be. It's wider and clocked higher than either one of those. This helps bandwidth.
2. It's full duplex and can send and receive at the same time. This will help latency, as the CPU or northbridge doesn't have to wait for the other to finish sending before they can do their own thing. It also doubles the overall bandwidth.
3. Communications, while still having to leave the cores to get routed, don't have to leave the socket until they're sent directly to the destination PHYs. This saves power as the environment between the cores and the communications hardware is always the same, and thus can be far more tightly tuned, resulting in lower latency and higher bandwidth.
There's also the rumor currently going around that AMD has decoupled the IF/Mem clocks. If true, AMD (or the end-user) could really clock the Core-I/O links to much higher values further helping bandwidth and latency. It also helps that nowadays cores have access to much larger cache pools than what was available back in the day.
Will there be a penalty for going the non-integrated route with the I/O die? Most likely. But the bonuses for doing it this way in AMD's eyes outweigh the downsides, and so they're doing it.