dec
Let me put it another way.
Each CPU has 128 PCIe lanes, right? Now, either all of those 128 are the same or 64 of those are multi-purpose, can be configured that is.
In a 1S system, they're all regular PCIe lanes. In a 2S system, 64 are used for PCIe devices, the other 64 are reserved for Infinity Fabric. Question: are only 64 of the total configurable as either PCIe or IF (or anything else for that matter)?
Yes. Like the PCH and other examples that preceded where there is a configuration switch PCI-e or USB/SATA/Ethernet/etc. it is typically only a two way switch. It is cheap to implement in terms of transistor budget.
It is like a two input , "a/b" keyboard-video-mouse switch. You can hook two computers to a single set of a keyboard monitor and mouse. Turn the switch to setting 'a' and can talk to one computer. Turn the switch to setting 'b' and can talk to the other computer. The switch allows to share the same resources. The shared resources in this case are just the wires ( more so like the cables coming out of the KVM switch to the keyboard/video/mouse).
This also isn't like a KVM switch in that it varies over time. Once you initially build the system it is a pragmatically down switching. ( If moved the CPU to another box with different wiring and firmware it would be set different, but is typically an extremely rare event. The frequency is totally different. )
Since the CPU is made of similar blocks, I'd imagine all lanes are equal, unless each block has configurable lanes of it's own, that when assembled together will be available as inter-CPU connection in the form of IF.
No. In order to make it so that you could flip them to whatever on a laundry liist of 2-4 things you'd need some far more complicated cross bar switch. That is a spectacular waste of space and transistors if the switch is set
once on configuration and pragmatically never changed once the CPU is coupled to the wired logic board.
DMI is also another example of PCIe lanes repurposed.
No, the DMI aren't generally available. You can't put other stuff on those lanes and have a viable system. DMI also has a slightly different protocol and has slightly different physical constraints since it is always a single point-to-point link.
" The DMI 3.0 protocol is ... essentially upgrading DMI from PCIe 2 to PCIe 3, but requires the motherboard traces between the CPU and chipset to be shorter (7 inches rather than 8 inches) in order to maintain signal speed and integrity. ..."
http://www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation/3
'essentially' and 'is the same' are not interchangeable. As this gets cranked up another level the distance will probably shrink again.
[doublepost=1495127194][/doublepost]
People moaned and groaned about the price raise on the MacBook pro but Apple's financial results show that it sold well. Obviously a $200 watch or a $400 tablet have different price sensitivities than a $2000 laptop.
The rumors are now tracking that Apple is going to speed bump these laptops to Gen 7 ( Skylake) after a little more than 6 months. So was it value customers were buying in that quarter's window or just a non-comatose laptop? I'm talking longer term pricing power, not about the narrow subset of folks who throw rush out and buy anything new.
I disagree. If AMD only offers a competitive advantage for a small portion of Apple's lineup, I don't think Apple would go to the trouble of adding support for AMD's chips and platform in macOS, especially if those benefits would only be seen in their "low end" products like the mac mini or iMac. Its much easier to stick with Intel even if AMD offers a slight advantage across some products.
You are presuming that AMD's Vega based APU only have a slight advantage over Intel's Iris/Iris Pro set ups. I'm not so sure that is going to be so 'slight'. It depends upon just how affordable the HBMv2 they can put in package is. If it is Vega's 1GB cache versus Iris Pro's 128MB cache, then I suspect the difference won't be slight on alot of workloads.
Intel caught up to and may have pasted AMD a bit on their iGPU competition. However, the combo of Zen and Vega may be throwing Intel for a loop. They have caught up to AMD iPGU by throwing the vast majority of the transistor budget toward iGPUs part of the die. Zen puts pressure to reallocate back to x86 core transistor spend (e.g., Coffee Lake going to 6 cores), but Intel is in a "rob Peter to pay Paul' mode at this point because would need to shrink the iGPU to do that ( there are no GT3 or GT4 options on table for Coffee Lake).
Apple is probably going to want both if they can get them. Same up tempo x86 core count threat is looming for the Mini and iMac and "Retina" screens for everybody is has not let up any pressure on the iGPUs. The 21.5 and Mini (with screens Apple wants to sell) has also a pain point. If AMD fills that pain point better they should go.
In short, it took AMD a ridiculous amount of time but they are pretty close to getting the dominance in the CPU+iPGU package market they wanted to generate 5-6 years ago.
If AMD Vega APU is cache-less then Intel probably is still in the game. but at this point I suspect AMD has the substantially better cache tech. At least for next year or so.
Sure, but the TB chip/controller is sold by Intel. Whats to stop them from refusing to sell it to someone implementing it on an AMD platform.
Intel refuse to make more money. Sure they could do that. Being a jack-ass is not going to help TB long term. The vast majority of the system vendors don't like Intel as a sole source supplier on this tech. If Intel is a "benevolent dictator' then folks will gradually go along. If Intel cranks up monopolistic tactics a cheaper USB 3.1 Gen 2 + DisplayPort switch is going to come along. ( there is a looming PCI-e Alternate Mode looming in background too. It has been a mentioned possibility for the next addition to an official spec update. )
More likely it is AMD dragging their feet because it does mean more money for them. Do you see Intel Ethernet controllers on AMD reference boards? Nope.
I think this is confirmed by the fact that no TB3 is found on any high end Ryzen motherboards despite demand for it from PC users.
AMD Ryzen based boards are having BIOS/Firmware problems with just their own list of approved stuff. Why would anyone add the complexity of more sophisticated firmware on top of that. They are woking out kinks in memory compatibility and other stuff. Open question as to whether folks are going to buy in large enough numbers.
Throw on top that all the Ryzen offerings so far have no iGPU. Standard PCI-e slot GPU cards and thunderbolt has a huge impedance mismatch. All the more so with the "Rube Golderberg" solution that the general box with slot market can settled on with funky loop-back cables.
When there is a large number of Ryzen APUs and no Thunderbolt then may have something substantive. However, at the moment this is not even remotely confirmed. The primary market for Ryzen 5 and 7 is affordable, off-the-shelf add-in-card GPU buyers. Even on the Intel side you won't see much TB present in those systems either.
