Nvidia Maxwell a bad choice for VR?

The problem is not with compute shaders, but with async compute, which is compute that is not for rendering. This is meant for tasks that do not need to finish at a particular time, and can even miss frames without causing a problem. Common examples are streaming texture data in the background and updating global illumination data. Any operation that involved drawing triangles cannot be async, by design.

So, it is not suitable for anything needing low latency. It is not something you want to use as a part of the render pipeline, since the whole point is that it is independent of the render pipeline, gaining flexibility by taking away timing guarantees and using the GPU's idle time to do other work.

VR requires you to do all your rendering work right now as fast as possible, and the primary rendering pipeline is going to be heavily utilized. Doing lots of async work is probably not a big gain, especially when things like realtime GI is not something we can afford in VR yet realistically.

Where async would be good for VR is the streaming case, where we could load and process new data onto the GPU without any framerate hitches. That would have been nice to have. Still, async timewarp will prevent loading from causing stutter most of the time. Async timewarp has nothing to do with async compute, because like I said you cant control the timing of async compute it would miss vsync constantly if you tried to use it for that.

Nvidia also still supports async direct memory access, so actually transferring data from CPU to GPU does not block the GPU. They use this for VR SLI to copy the frame buffer from one GPU to the other. The case where you lose async performance is when you need to have the GPU do a lot of processing on the uploaded data, like perhaps for virtual texturing where you need to copy data around a lot inside the GPU after it is uploaded, that would benefit from async compute.

There are also probably other techniques nobody has thought of yet that wont perform as well on maxwell compared to AMD. But the stuff people are doing now will work fine, and the massive draw call boost of DX12 will still work fine.

"owenwp" wrote:

Nvidia also still supports async direct memory access, so actually transferring data from CPU to GPU does not block the GPU. They use this for VR SLI to copy the frame buffer from one GPU to the other. The case where you lose async performance is when you need to have the GPU do a lot of processing on the uploaded data, like perhaps for virtual texturing where you need to copy data around a lot inside the GPU after it is uploaded, that would benefit from async compute.

There are also probably other techniques nobody has thought of yet that wont perform as well on maxwell compared to AMD. But the stuff people are doing now will work fine, and the massive draw call boost of DX12 will still work fine.

They stated to have developed memory to actually be on the GPU. I watched a video way back of a company saying they are 3-4+ generations ahead for development.

When VR gets huge and the funds becomes massive because to have a great VR experience you have to have lots of processing power then nvidia will release.

Well whatever direction VR is going it looks to be very promising. : )

There's many cool stuff in the world that had unfortunately died because no one of power/talent/vision would make it back into reality. Cuisinart bought the foreman grill so that got resurrected. I still see many stuff that marketing has died because the flow is good enough so everyone "new" would never know about it unless by accident like the Sonicare toothbrush.