NVLink vs. SLI

[Prima.Vera]

As title suggests, what are the biggest differences between NVlink vs. SLI, but more import advantages using one over the other.
Last but not least, are ALL the games compatible with the new NVLink, or they need special implementation just like for SLI?
Thanks!

 

[Mussels]

same software, same motherboard support, just need a new bridge on the new cards

 

[Prima.Vera]

I thought the major thing, except the huge bandwidth increase, is that with 2 cards you are actually doubling your VRAM and performance too, and you can go up to 6 cards...

 

[Mussels]

I thought the major thing, except the huge bandwidth increase, is that with 2 cards you are actually doubling your VRAM and performance too, and you can go up to 6 cards...

I havent seen anything about that yet, probably wont find out the finer details of that stuff until reviews are done

 

[newtekie1]

I thought the major thing, except the huge bandwidth increase, is that with 2 cards you are actually doubling your VRAM and performance too, and you can go up to 6 cards...

I don't believe NVLInk is fast enough to allow VRAM sharing. The server implementation of it is only 150GB/s duplex(s0 300GB/s total but only 150GB/s in each direction). This wouldn't be fast enough to allow VRAM sharing between the cards. And that is assuming the NVLink that we get in the consumer space is the same as what they are using on the servers.

I'm betting it is a scaled down version for the consumer cards. HB SLI was only 2GB/s, and I'm guessing they are only switching to NVLink to support higher resolution that HB SLI couldn't handle, and maybe to go back to properly allowing more than 2 cards.

 

[Upgrayedd]

I don't believe NVLInk is fast enough to allow VRAM sharing. The server implementation of it is only 150GB/s duplex(s0 300GB/s total but only 150GB/s in each direction). This wouldn't be fast enough to allow VRAM sharing between the cards. And that is assuming the NVLink that we get in the consumer space is the same as what they are using on the servers.

I'm betting it is a scaled down version for the consumer cards. HB SLI was only 2GB/s, and I'm guessing they are only switching to NVLink to support higher resolution that HB SLI couldn't handle, and maybe to go back to properly allowing more than 2 cards.

I haven't seen anything but what about an updated NVLink? Like the LED bridge and HB Bridge are to the standard SLI bridge.

 

[R-T-B]

I haven't seen anything but what about an updated NVLink? Like the LED bridge and HB Bridge are to the standard SLI bridge.

We'd most likely have seen it in server space first. They get all the cool stuff first, we get the leftovers.

 

[londiste]

SLI bridges have bandwidth of 1GB/s (normal bridge) and 2GB/s (HB bridge).

NVLink on Turing cards is 2 links. Each link can do 25 GB/s both ways. On Turing, total is 50GB/s one way and 100GB/s total (with the last one being the marketing number).
Consumer cards - that is, Geforce RTX - have one link disabled, 25GB/s one way and 50GB/s total.
That also means it can only be one link - 2 cards connected together.

Memory bandwidth is 616 GB/s or 448 GB/s for GTX2080 and GTX2080Ti, there is a huge difference there. Memory sharing for real-time/gaming is not a viable thing.
GDDR5 version of GTX1030 has memory bandwidth of 48 GB/s if you want a comparison

As far as SLI or otherwise cooperation implementation goes, that is entirely up to Nvidia. I would bet for consumer it will just be a slightly better SLI. Even that 10-fold increase in the available bandwidth will not change the inherent issues with current rendering techniques and multi-GPU.