Pyro as Bill said:
The Akitio Thunder 2 is made for Thunderbolt 2 but in the top two cases they're only using a Thunderbolt 1 connection. The PE4 being used with the HD7970 is using the expresscard slot. The lack of lanes means 4gbps is their max. Some people only had 2gbps to work with and still saw huge improvements.
I get it. Even though USB 3.1/C and a Thunderbolt 1 connection are both 10gbps, USB doesn't have a direct connection to PCI-E like TB/EC/miniPCIe/m.2 so latency hinders it. Ping is more important than speed.
|
The Switch doesn't have USB 3.1 though, it has a slower USB 3.0 hub which is operating at USB 2.0 speeds. USB is still higher latency than even thunderbolt 1.
Latency is a *massive* factor in synchronizing multiple graphics processors, it was the single biggest issue with AMD's frame pacing woes going back a few years ago... To the point where they abolished the Crossfire bridge and threw everything onto the PCI-E bus.
The limited bandwidth the Switch's USB does have needs to reserve a massive chunk of bandwidth for the docks hub, video, handshaking, audio, networking, external storage and so on as well, it cannot all be dedicated to Multi-GPU.
GPU's can technically get away with slower interface speeds because of the amount of data they load into the GPU's dedicated DRAM before running the game, meaning load times are slower. - Where you see a massive impact in performance is when you start to stream assets into the GPU's DRAM... And tons of games asset stream these days, it was one of the big focal points pushed during the 7th gen.
Back in circa~ 2006 I was doing testing on a PCI Radeon x1300 doing modding for Oblivion, once the game had loaded all the assets into the game, the framerates were fantastic, but once the game started to stream the assets for the next cell into the GPU's memory, performance would tank. - The issue continued to exist even when I decided to ditch that GPU and opt for a PCI Radeon 4350.
Obviously, the PCI slot is an extreme case as it only has 133MB/s.
But any PC enthusiast can tell you that running a modern GPU at PCI-E 1x speeds is not going to be a good time, your minimum framerates *will* suffer.
It also does mean that you will essentially need a second switch in the Dock, complete with DRAM, in order to cut down on the amount of data being transferred.
Pyro as Bill said:
Theoretically, all Nintendo would have to do is put the connection in and a 3rd party could do the rest (or most of it)?
A 1030 is 50 flops faster than a 980 but 50 flops slower than a 1080, correct?
|
Indeed. Theoretically Nintendo just needs to revise the Switch and add a PCI-E based connection to the Switch, nVidia already has the technology available to make it seamless... Or Nintendo can leave it up to the developers.
****
As for Flops, they are theoretical numbers, not real world... And in the context of discussion it is entirely irrelevant anyway, nor are flops representative of a systems capabilities anyway.
