PCIe lanes for peripherals

[daanav]

Hey everyone,

I have a question regarding PCIe Lanes. I am planning to have a configuration with one GPU, one M.2, one SSD and a network card. If I get 8700K, the would the peripherals affect the 16 Lanes from the 8700K to the GPU?

Thanks in advance.

[Crunchy Dragon]

The network card might, depending on if it runs on chipset PCIe lanes or CPU PCIe lanes.

 

Either way, a GPU can go down to PCIe x4 speed before you'll start noticing bandwidth issues.

[daanav]

Hey, thanks for the reply.

I want the GPU to always have 16x as it will be used for neural network calculations.

[Crunchy Dragon]

Just now, daanav said:

Hey, thanks for the reply.

I want the GPU to always have 16x as it will be used for neural network calculations.

I don't think the GPU will be affected by a network card, usually only with higher bandwidth things such as another GPU does your system have to split or redistribute PCIe lanes.

 

I believe the network card will run off of chipset lanes.

[daanav]

Does this apply for both wireless cards and Ethernet cards?

Because when I look at wireless cards, it says 1x PCIe is needed. Do they mean PCH here?

[porina]

Consult your mobo manual as it should tell you which lanes go where. Generally on consumer level systems (not HEDT or server) two x16 physical slots go to the CPU and run as either 16x+0x or 8x+8x depending on which are used. Other slots run off the chipset so wont affect it, but may be shared with other mobo features.

[mariushm]

There's 16 lanes coming directly from CPU (on most processors) to that first pci-e x16 slot ... those lanes are reserved for video cards.

 

The other pci-e slots on your motherboard are connected to the chipset - the chipset creates a number of pci-e lanes which are then grouped in chunks of x1 , x4, x8 or x16 and assigned to slots.  The motherboard bios can choose to restrict a slot to fewer lanes to rewire some lanes to other places, like m.2 connectors

 

On some motherboards (depending on chipset) it's possible to have those 16 lanes coming from cpu split into two x8 and on those motherboards you have two pci-e x16 slots for video cards (for SLI, Crossfire etc) - those x16 slots become electrically x8 unless the motherboard has special (expensive) chips that convert those x16 lanes into 2 x16 lanes.

 

Anyway, with those motherboards that have 2 x16 slots for video cards, some will allow to insert any kind of pci-e card into that second pci-e slot (and the two slots become pci-e x8 slots) but some won't .

 

So yeah ... anyway... your configuration will work just fine. Your video card will be in the pci-e x16 slot and always have x16 lanes of performance, your m.2 will be from chipset, the other pci-e slots will be from chipset.

Depending on motherboard, one of the pci-e slots may become unusable or restricted to x1 (for example one pci-e x4 slot becomes x1 electrically if you use m.2 drives)

[daanav]

@mariushm Thank you for the explanation. It helped a lot.

[JoostinOnline]

57 minutes ago, daanav said:

Hey, thanks for the reply.

I want the GPU to always have 16x as it will be used for neural network calculations.

I doubt it will ever go that high. We'll see what happens with the new cards, but no existing card uses more than 8 PCIe 3.0 lanes (the Titan V).

[daanav]

@JoostinOnline Are you sure that the Titan V will also not max out the 16x PCIe 3.0? I thought it is in gaming that the GPU will not max out but I think in machine learning it will surely max out the 16x Lanes as very high data transfer rates will be required.

[JoostinOnline]

7 minutes ago, daanav said:

@JoostinOnline Are you sure that the Titan V will also not max out the 16x PCIe 3.0? I thought it is in gaming that the GPU will not max out but I think in machine learning it will surely max out the 16x Lanes as very high data transfer rates will be required.

Not that I know of. It was a pretty big deal when they were able to approach the 3.0 bandwidth limit, but they were only able to do it with two Titan V's in SLI, and in x8 mode.

[daanav]

@JoostinOnline They tested in Gaming environment or calculation environment? Because for example if I have a batch size of more than 8Gb which is possible in deep learning, then it should reach a bottleneck if running at 8x. Right?

[JoostinOnline]

Just now, daanav said:

@JoostinOnline They tested in Gaming environment or calculation environment? Because for example if I have a batch size of more than 8Gb which is possible in deep learning, then it should reach a bottleneck if running at 8x. Right?

It was in rendering. I don't know anything about deep learning so I probably shouldn't be giving advice on them.

[daanav]

@JoostinOnline Ok. Thanks for answering the question. It was helpful. And thanks to everyone else also. I will try to find some tests done by machine learning engineers. 

 

But now I know more about PCIe lanes. So can dig further.