Z68 - NF200 Chip - Technical Question

[dazzerd]

OK have been reading up how the NF200 chip works on Z68 (P55, P67) and I still can't get my head around it.

All the above chipsets have 16 pci lanes on the cpu and 8 lanes through the southbridge.

Boards excluding the NF200 chip either use the full 16 lanes on a set up but would split them into 8 x 8 for sli/crossfire. All basic stuff.

The NF200 claims to add 32 lanes, my understanding of how it does this is by multiplexing the cpu lanes, basically multiplying them by two. This gives boards with the NF200 the ability to run dual card set up with full 16 x 16.

However the GPU must communicate with the cpu to output onto a display. The CPU on these chipsets can only connect with any gpu devices with a maximum of 16gb's as the lanes are actually on the chip. Therefore how can two full 16 x 16 cards (with max throughput of 32gbs) communicate at full bandwidth when the connections has to be physically limited by the links available on the cpu?

I;ve read lots of explantions but most are either confused or unconvincing. Surely any devices connect to the cpu in these cases has to be limited to the links on the cpu?

 

[EVGA_Nick]

The main advantage on platforms with only 16 PCI-E lanes is the ability to have more than 2 full sized PCI-E slots (not counting the one connected to the PCH). Whether this is for 3-way SLI, a RAID card, a PCI-E SSD, etc. that would be the main benefit. You can run 2 cards at x16 on our boards but it'd be better to run a direct x8 connection to the CPU on one of them in my opinion After all, the 2x16 lanes are only x16 up to the nf200 chip. This would go the same for a board claiming Gen 3 support but may have lane multiplexers which may not support it so only some configurations may support Gen 3.

Here is an example of how EVGA has implemented it into P55/P67/Z68 F T W versions. (P55 used as the example)

 

[Stone22uk]

The main advantage on platforms with only 16 PCI-E lanes is the ability to have more than 2 full sized PCI-E slots (not counting the one connected to the PCH). Whether this is for 3-way SLI, a RAID card, a PCI-E SSD, etc. that would be the main benefit. You can run 2 cards at x16 on our boards but it'd be better to run a direct x8 connection to the CPU on one of them in my opinion After all, the 2x16 lanes are only x16 up to the nf200 chip. This would go the same for a board claiming Gen 3 support but may have lane multiplexers which may not support it so only some configurations may support Gen 3.

Here is an example of how EVGA has implemented it into P55/P67/Z68 F T W versions. (P55 used as the example)

Wow top answer

 

[dazzerd]

Yes thanks for that clarification and honesty.

However I'm still confused how all this extra bandwidth is negotiated over the 16 slots available via the cpu. My interpretation of your diagram may be wrong but by the sounds of it you are saying that depending on card configuration you have 40 pci-e lanes available for add in cards, gpu, raid or whatever. 8 of these direct to cpu the other 8 retained for the link to the nf200, how then can all 32 slots available via the nf200 if fully populated communicate with the cpu when there are only 8 lanes to do this?

Sorry for the questions I'm just interested

 

[Jokester]

The benefit I guess is that some of the bandwidth is used purely from one GPU to the NF200 and then to the other GPU, so it doesn't waste any of the bandwidth back to the CPU that would happen if you didn't have the NF200?

 

[mmj_uk]

nvm

 

[dazzerd]

The only thing which comes to mind to me is if the nf200 acts like some kind of very very fast traffic controller. So each card linked to the nf200 only has a certain time to communicate with the cpu before the lanes are made available to the next device and so on and so forth. Would also explain the extra lag over the nf200.

Interesting though how they are marketed as making available full bandwidth to the gpu's when common sense seems to suggest that that cannot be the case

Nice boards though Nick; fancy sending me z68 F T W so I can test my theory

 

[EVGA_Nick]

So each card linked to the nf200 only has a certain time to communicate with the cpu before the lanes are made available to the next device and so on and so forth. Would also explain the extra lag over the nf200.

nf200 can output 32 pci-e 2.0 lanes. It doesn't matter what you are connecting as an input, whether it be x16 or x8. Lanes are not 'put on hold' as you suggest. They still operate just like any other pci-e lane.

Interesting though how they are marketed as making available full bandwidth to the gpu's when common sense seems to suggest that that cannot be the case

Well, the cards can run at x16 PCI-E 2.0 and can also communicate with each other via nf200 at this speed. In my opinion, the added latency is made a much bigger deal than it actually is. I would worry more about finding a board/configuration that matches your needs