![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
Got the new Gen 3 splitter today and it's quite a bit different from the previous one that was Gen 2. Hopefully this one will work with the 970 GTX.
Got the new Gen 3 splitter today and it's quite a bit different from the previous one that was Gen 2. Hopefully this one will work with the 970 GTX.
Man, this is awesome. Come on PCIe 3.0!
Got the new Gen 3 splitter today and it's quite a bit different from the previous one that was Gen 2. Hopefully this one will work with the 970 GTX.
SLI throws a wobbler if you try and run it without all the cards on 8x or above. Which is a bit silly; 4x 3.0 is effectively the same speed as 8x 2.0, and while SLI will work with all cards on 8x 2.0, it won't with them on 4x 3.0.
175w TDP x4 would need quite a beefy psu and cooling surface area
That is old DirectX11 thinking ;-).
I'm not sure where this idea comes from. DX12 is not magic, it just takes the job of partitioning out jobs between GPUs from the driver developer and puts that task in the hands of the engine and game developer. If you do not have a workload that scales well to multiple GPUS, it will not scale well regardless of API.
This is a MASSIVE oversimplification. For one thing DX11 already HAS multithreading implemented!:
Some benefit will come from job dispatch having reduced overhead, but this will mainly be a benefit in scenarios which are currently CPU bottlenecked not the case with most games) or where driver optimisation has lagged behind (e.g. recent AMD drivers).
No, the reason scaling is at best "OK" with 2 cards and progressively worse with more cards is because, while GPU rendering tasks are internally embarrassingly parallel (particularly pixel-level operations), GPU jobs are not. Not only are they sequential (you need to know where your polygons will be before you can determine if they are visible, you need to know if they are visible before you can shade them, etc) but also interdependent (e.g. the leaf geometry of a procedurally generated tree is dependant on the position of the tree 'base') meaning that data just be shared between GPUs, and in the correct order, or rendering will break.
It also requires all the work that is currently done by Nvidia (and AMD) to distribute jobs between GPUs to be done by the engine or game developers. This is not something that they necessarily have any experience with, even if they have been working very closely with GPU vendors. Threading is a Hard Problem, and for many tasks will result in minimal gain even with perfect optimisation. Developers are ALREADY pushing threading as hard as they can to take advantage of modern quad-core CPUs (and to get the most out of consoles), and the scaling benefits are FAR from dramatic.
HBM and HBM2 are going to do nothing for data sharing between cards, as the bottleneck is the shared card bus. NVLink remains to be seen: all Nvidia's presentation materials have focussed on NVLink with PowerPC for HPC applications, though a handful of slides show NVLink in use as an auxiliary GPU-GPU (i.e. NOT CPU-GPU) link for x86 systems. And what we've yet to see is the physical form-factor of NVLink; it may be something like the SLI bridge where it could be compatible with existing PCIe specs for consumer devices, or it could have such rigid electrical specification that it needs an auxilliary slot and routing through the motherboard (in which case the chances of it turning up in consumer x86 boards is slim-to-none).
This is not the case. DX12 Implicit Multiadapter is NOT the same as DX11 SLI/crossfire. SLI/crossfire is doing some degree of per-card job dispatching, whereas Implicit Multiadapter is a much more basic mirroring of memory and jobs between cards for AFR.
This is obviously passive splitter but I got confused with so many splitters which Slime is juggling. What I would like to know if there is method to do the splitting without special BIOS. Equivalent question is how splitting is done with various expander boxes where just special cable is connected to a PCIE slot.
In theory:
