re: PC Discussion - Gaming, Performance and Enthusiasts

Insane

0.27/GB

What's the latest on DX12, if it does support VRAM stacking that will be huge for my SLI configuration.

quote:

---

What's the latest on DX12, if it does support VRAM stacking that will be huge for my SLI configuration.

---

Lots of awesome promising things that DX12 will allow. Hell, it even enables the use of mixed configurations of GPUs because DX12 treats all of your GPUs as one single GPU (e.g., using a 290 and a 970 in some weird SLI/crossfire hybrid and maybe even throwing some Intel iGPU power in the mix).

The theoretical possibilities sound great. The reality will probably be disappointing when you consider that all of these possibilities still have to be implemented by game developers and at the driver level.

Someone on the Linustechtips forum brought up a good point about a month ago -- the problem with just generically stacked or pooled vram (i.e., treating SLI as one big GPU and the vRAM as one pool of memory communicating with the one big GPU) is bandwidth limitations. As it is now, textures are loaded into vRAM via PCIe bus. The bandwidth of even a PCIe 3.0 x16 slot is around 16GB/s -- that's as fast as the cards could possibly transmit data between each other. The data, once in vram, has to communicate with the GPU exponentially faster than what current PCIe bandwidth allows. The GTX 970's memory bandwidth is 224 GB/s, for example, so GPU A can access VRAM A at 224 GB/s. If GPU B needed to access VRAM A to render a frame, it would currently have to do so across PCIe, and the bottleneck would be horrendous. Nvidia is working on a solution to this called NVLink, but I think it's still being tested in GPU-powered supercomputers. I'm sure it will eventually hit the consumer market, but it's not something you'll just be able to tack onto your current hardware.

With current hardware limitations, for pooled vRAM in a standard crossfire/SLI setup to work, developers would specifically have to pay attention to the way their game is rendered with multiple GPUs -- probably via split-frame rendering, where one GPU renders the top half of a frame, and the other the bottom half, so the texture data in each card's vram doesn't have to be mirrored like it does with alternate frame rendering, and the GPUs won't need to access each other's data... I think... there could be other creative ways to solve the problem, but we'd still be at the mercy of game developers, and that's never a good thing.

I read earlier about what you just described. Sound like by the time it's figured out I will be ready to upgrade to new cards anyway.

SLI 970s way better than a single 980 Ti. Even beats a single Titan. You won't find a single GPU solution beating the 970s for a while.

quote:

---

SLI 970s way better than a single 980 Ti. Even beats a single Titan. You won't find a single GPU solution beating the 970s for a while.

---

Gotcha, thanks.

I emailed them thanks for the link, we will see what and how fast they respond. What would you suggest regarding adding water to the cards, don't do it or if so what hardware if you don't mind. Would a separate loop just for the cards be better than adding to the swiftech?

I will hold off then, I am going to pull and fan from an old computer and add inflow on the open bottom spot. I do need to put some space between the cards, the SLI bridge will let me put one more slot between them

Thanks again for your help.