cekim wrote:
It is a little strange to have 3 generations of chips with a 4th (broadwell) hiding in the middle that are all strikingly similar and on top of that to lose total channels/bandwidth with the latest.
Not really strange. The PCIe 3.0 lane controllers are built into the processor die and they can't be moved off die because the whole point is to have their bandwidth piped directly into the processor without any bottlenecks. Intel copies the same block of PCIe Uncore circuitry onto all their processor designs, having a "standardized" design makes all their PCI SIG chores much easier because the "standard" is - as always - impelled more by actual device intercompatibility than by arcane technical datasheets.
And Intel makes their real money selling server tech. Performance is important but "standardized" compatibility (and reliability, and cost) are far more important. They could (and eventually will) design processors with more integrated PCIe lane controllers. The server folks tend to fill up their PCIe slots with fast storage and workstation cards, more lanes are always better but for now it's more economical to add more processors and blades to the racks than it is to upgrade the computing density built into them, especially since more Uncore on the die must mean less of something else on the die (or worse, it must mean larger and more expensive silicon). It turns out that the PCIe controller block isn't easy to shrink down on smaller fab lithography, just one of those things which has to be a certain size to work properly - we can count on Intel to figure it out (at some astronomical cost, no doubt) but we're apparently gonna be stuck with 40 lanes for long while to come.
Enthusiasts and powergamers want x16/x16/x16/x16 multi-GPU setups. Many of us also wouldn't mind maybe 8 more lanes to run superfast boot drives. Companies like Intel and Asus are more than willing to accommodate our desire to pay top dollar for top tech. But we're not even a tiny drop in the bucket compared to what those big server fleets are worth, we're good for "researching" overclock-related stuff but the IT platforms are where all the demand (and money) for uber tech all comes from.
I'm still expecting (hoping) that next-gen LGA2011-3 processors will pack PCIe 3.0 controllers capable of running 48 or 56 or 64 lanes. But it's just not going to happen until Intel perceives a lucrative demand for more lanes in the server arena (or at least the opportunity to create such a demand). AMD favours chipset-based PCIe 3.0 controllers so Intel really has no competition to worry about, they already dominate the market, and they'll probably milk their advantage for as long as they can before needing to push this area of technology forward.
"All opinions are not equal. Some are a very great deal more robust, sophisticated and well supported in logic and argument than others." - Douglas Adams
[/Korth]
