AMD Vega 20 Supports PCIe 4.0

[jimmysmitty]

That and the ability to even push some to PCIe 4.0 x4 to make shorter cards for HTPCs and such.

 

[Gillerer]

GPUs can - and probably will - still be x16 or x8, but if motherboards don't absolutely need to allocate at least x8 to each GPU, there are more possibilities of innovative PCIe lane allotments.

I could see a mainstream motherboard with four PCIe 4.0 x16@x4 slots with enough bandwidth in each to comfortably run a GPU. (You'd want to have switches to allow unpopulated slots' lanes used by the populated ones, obviously, like current mainstream x16/x0 or x8/x8 system.)

Or having the option of splitting a x8 GPU slot so it only uses x4 lanes and the other x4 are given to an additional M.2 slot when it's populated.

 

[jimmysmitty]

The issue with x4 is power delivery. x8 delivers more power and on high end GPUs you need the power from the slot plus the additional. At least thats how most GPUs are designed, to pull the max from the slot and the rest from 6/8 pin so I don't see high end dropping below x8 but mid range could drop to x4 easily.

Of course another option is adding PCIe lanes to the CPU as most GPU lanes are pulled directly from the CPU or leave the CPU with that and add more to the chipset. Intel currently has this setup where they have at least x16 from the CPU dedicated to a x16 slot and then around 24 on the chipset that are allocated to everything else.

 

[Gillerer]

Adding lots more PCIe lanes to mainstream and low-end CPUs is putting resources (and money) where the vast majority of people will never use them, so I don't see that happening. *Maybe* Intel will add 4 lanes to allow direct M.2 connection like AMD. Then again, with a PCIe 4.0 x4 level DMI, the bottleneck between CPU and chipset will be much less of an issue in even those most demanding scenarios, so maybe not.

*

I'm sure the PCI-SIG have thought through all the usage scenarios, so maybe they'll bring the ability to power 75 W high power devices down to x4, too?

If not, even now the x8 slots on motherboards are physically x16, and many motherboards even have x16@x4 slots. If the slot is x16 size physically, it can be engineered to supply x16 levels of slot power. It's all about the negotiation during initialization between the GPU and the motherboard. I don't know whether the x4 data connection itself imposes any limits to power - it's just that the slot is not *meant* (as per standard) to be able to supply a PCIe 3.0 x4 device more than 25 W...

 

[jimmysmitty]

I don't see any issue with M.2 on the chipset vs CPU. The drive itself will be more of a bottleneck than that will ever be as the NAND hits limits. Plus if Intel has its way they will push the crap out of XPoint and NVDIMMs instead which would drop the need for M.2 links to the CPU as direct memory connections are faster by far.

 

[Gillerer]

I don't think it's that much that bottlenecking is currently a problem, but that looking forward, doubling the DMI speed will make sure it won't be a bottleneck in the near future, either, other than very special and rare setups.

Even now the chipset connection can be saturated by lots of bandwidth-hungry devices, like two higher end NVMe drives, or one such drive together with an array of SATA SSDs and maybe a 10 Gb network card. Imagine if NVMe drives get 1.5 to 2 times as fast as they currently are and more and more motherboards implement built-in 10 Gb NIC (through the chipset PCIe lanes).