Nvidia kills 3x and 4x SLI with Pascal *spawn*

It was interesting how he just mentioned 2xSLI 1080 beating 4x980ti.
Interesting because it was just confirmed today that the 1080 is limited to only 2xSLI, so no 3 or 4 card combos.
Cheers

Why would Nvidia want to kill 3 and 4 way SLI? I dont get it.
 
If it's not for a particular technical reason, they will have data for how much the 3/4-way SLI market is worth and how much extra work it requires of their driver team and perhaps concluded it wasn't worth supporting.
 
Maybe Nvidia rather have consumers buy 2x 1080 than 3x or 4x of the cheaper 1070?

Anyways, still not sure why so much Nvidia talk is in an AMD thread.
 
Are they only killing the old style transparent SLI only? DX12 multi adapter with devs hands on approach should still work with more than 2 cards? Or they limiting their driver to only support 2 cards?
 
Could it be that they require NVLINK connection between the chips for SLI with Pascal series(to get rid of problems related to slow bandwidth between the chips),
and the 1080 chip only has only one NVLINK port?
 
Word is that in order to double the bridge bandwidth, they've decided to use both SLI fingers on the cards in parallel. So, because both fingers are used for 2 cards, you can't scale beyond that. Additionally, the 1070 will feature the older style SLI bridge and should theoretically be able to scale beyond 2 cards.
 
It was interesting how he just mentioned 2xSLI 1080 beating 4x980ti.
Interesting because it was just confirmed today that the 1080 is limited to only 2xSLI, so no 3 or 4 card combos.
Cheers
Ah blast, this was originally posted in the other thread to confirm the guy seems to have good information before it is actually announced, not to really bring up this actual news lol.

I can see a couple of reasons though why they would do this IF he is correct about its performance.
While it is known now there is only 2x SLI this is now what they call HB, what has not be shown publically until reviews is how well is scales, if performance is as good as 4x then 2xHB SLI is enough IMO.
Part of this could be to protect FP32/FP16 professional market sector who are being prodded towards Tesla/NVLink or possibly Quadros (that may support higher than 2x HB SLI).

I guess we need to see the reviews and its results, but he was spot on just mentioning the 2X SLI for Pascal.
Cheers
 
Last edited:
Could it be that they require NVLINK connection between the chips for SLI with Pascal series(to get rid of problems related to slow bandwidth between the chips),
and the 1080 chip only has only one NVLINK port?
NVLINK is apparently present only on GP100. This is just same old SLI, but instead of daisychaining several cards, you're limited to 2 because it uses cards both SLI-links.
 
NVLINK is apparently present only on GP100. This is just same old SLI, but instead of daisychaining several cards, you're limited to 2 because it uses cards both SLI-links.
What I mean is it will appear on other Tesla cards not just the P100 (there will be a few lower spec models).
No idea about Quadro, but like I say they may also be protecting that in some way or that also has the same limitation to ensure Tesla sales are not cannibalised.
If (a bit if) they have managed to increase the SLI HB performance to match 4x standard SLI, that must be a consideration on people buying that instead of the professional cards if they could use SLI HB to do nodes of 4x.
One reason I could see it being disabled, of course it may be some kind of technical limitation but seems strange they only manage 2x with the HB when standard and even NVLink can do 4x node.
Cheers
 
Last edited:
What I mean is it will appear on other Tesla cards not just the P100 (there will be a few lower spec models).
No idea about Quadro, but like I say they may also be protecting that in some way or that also has the same limitation to ensure Tesla sales are not cannibalised.
If (a bit if) they have managed to increase the SLI HB performance to match 4x standard SLI, that must be a consideration on people buying that instead of the professional cards if they could use SLI HB to do nodes of 4x.
One reason I could see it being disabled, of course it may be some kind of technical limitation but seems strange they only manage 2x with the HB when standard and even NVLink can do 4x node.
Cheers
I'm not sure I follow.
SLI HB = old SLI, but instead of daisychaining via 2 links, 2 links are now dedicated between 2 cards.
The SLI HB has just as much total bandwidth as old SLI did, but it's now usable on just 2 cards
 
What I mean is it will appear on other Tesla cards not just the P100 (there will be a few lower spec models).
No idea about Quadro, but like I say they may also be protecting that in some way or that also has the same limitation to ensure Tesla sales are not cannibalised.
If (a bit if) they have managed to increase the SLI HB performance to match 4x standard SLI, that must be a consideration on people buying that instead of the professional cards if they could use SLI HB to do nodes of 4x.
One reason I could see it being disabled, of course it may be some kind of technical limitation but seems strange they only manage 2x with the HB when standard and even NVLink can do 4x node.
Cheers

Still only IBM with Nvlink motherboard chispet capable, or i have miss something ? .. As for the "new SLI" HB .. well.. i m not quite sure what you do mean with it. I sadly think this will not change many thing ( for be honest i dont think it will change anything on the SLI performance )

Cant really compare GP100 announcement with the feature on GP104 ( outside i have not a copy of the die of GP100 ); many things are not in the logic of the GP104 who is in the GP100 it seems.
 
Last edited:
I'm not sure I follow.
SLI HB = old SLI, but instead of daisychaining via 2 links, 2 links are now dedicated between 2 cards.
The SLI HB has just as much total bandwidth as old SLI did, but it's now usable on just 2 cards
The cards can be placed in SLI, Nvidia developed a new SLI HB-bridge (high bandwidth) that offers twice the bandwidth compared to the regular SLI bridges you guys already know.
http://www.guru3d.com/news-story/nvidia-announces-the-geforce-gtx-1070-and-1080.html
 
Guys,
my post came around because someone in the know suggested a possible rumour with Polaris and its timescale.
This guy also pointed out a little while ago that the 2x SLI HB of 1080 is as fast as 4x980ti in SLI.
It was only confirmed yesterday there is no legacy SLI on the 1080, it only supports the NVIDIA HB SLI and this is limited to 2 cards.

So the only solid news is that NVIDIA has stopped the standard SLI for 1080, however the guy who broke the news of its performance may also be in the know because he did that a fair bit before the news became official of the 2x HB SLI being max....
This suggests it is far from a standard SLI connection, but only a rumour and we will not know until someone leaks it properly from a review :)
Cheers
 
Maybe nvidia wants to actively prevent 3 and 4-way SLI from existing, because of VR.
Imagine trying to explain someone who spent $1800 on 3 graphics cards why one of said graphics card needs to be put away during VR.
 
exactly - nothing changed except that they now use both channels instead of one, the cards already had both channels before, but could only use 1 per connected card, now they use both
As I said there is no information released how it works; only assumptions in the press or the guy who said 2x HB sli is as fast as 4x 980ti even before it was known that it was limited to 2xsli.....
We will get a better understanding when we see detailed benchmarking down with 2x1080, until just speculation and a leak that may be true (only putting some weight behind this as they nailed 2xHB SLI).
Cheers
 
They said old sli bridges still works.... If they are not driver limited, you should still able to create a very unbalanced 3/4-SLI config.


They offer twice the bandwidth because you are forced to use two bridges at time ( ͡° ͜ʖ ͡°)
This is called marketing.
Did you get that information from NVIDIA or an AIB/site?
I would need to try and track them, but I have heard mixed signals from that only 2 is supported natively, to NVIDIA supports only 2 for best performance (meaning more can be added but performance will be worst so sounds like optimised for 2 HB).

You have to love NVIDIA's communication with this card, now this after the Founder Edition confusion :)

Cheers
 
Last edited:
Back
Top