Cuda 6

http://insidehpc.com/2013/11/14/cuda-6-release-simplifies-parallel-programming-unified-memory-drop-libraries/

oh yus

Unified Memory sounds like a dream! I hope Cycles can used that to avoid memory limits for the GPUs!

It looks like the new specs. will indeed help with the effort to continue the part of Cycles development that involves features running on the GPU.

However, this means that using the new instructions may require some people (like usual), to purchase new cards so they can take advantage of all of the latest technology, as it seems like GPU generations are getting deprecated in terms of them as rendering devices much faster than CPU generations (where Cycles is able to fully support even old chips from 6 or more years back).

With Nvidia’s flagship cards approaching 1K now, that means that you really need to save up the money if you want to get the latest and greatest out of GPU-tech, hopefully we can see Cycles become much faster on the CPU so we can diminish the entire need to keep up with GPU purchases.

Hi, Nvidia announce GTX 780Ti 12 GB for next month, they push GPU rendering really in last month.
12 GB is more than most user has as RAM.

http://www.hardwarepal.com/nvidia-announces-gtx780ti-12gb-gddr5-4k-4k-surround-gaming/#_

Cheers, mib.

I’m OK with GPU rendering, I’ve been using it for several productions during the last years and it’s very good. I don’t miss CPU rendering at all :wink: . Many of the memory limitations can be avoided by using layers/composition while the rendering it self is FAR faster in GPU than in CPU. Specially now with cards as Nvidia Titan and with Cuda 6 bringing us unified memory the future of the GPU rendering is SO bright :wink:

About the unified memory though, how much do you think performance would be affected if the scene had to borrow some of the CPU’s RAM to have enough memory for the scene, would you need to have some newer generation GPU slot type to ensure that there’s as little impact as possible?

If they use current tech, the hit will be pretty substantial. The speed increase rendering on the GPU is mainly due to the fact that all data is copied into vRAM so nothing has to be transferred over the bus after rendering starts. If memory is constantly being swapped between vRAM and system memory a lot of the benefits of GPU rendering will take a hit.

Then technology would first have to catch up in a way where there is a much faster and much wider memory bus that can shuttle data between the RAM chips and the GPU at blazing speeds (which would mean an entirely new socket type that would be added to motherboards).

If this is one of the things that will be needed for GPU rendering to truly replace CPU-rendering for most people, then I don’t anticipate this type of thing being in the hands of a wide audience for a few years yet (another part of it being able to expand the GPU cores to do more complex tasks without increasing heat and power output while maintaining performance).

Hi, Nvidia announce GTX 780Ti 12 GB for next month, they push GPU rendering really in last month.
12 GB is more than most user has as RAM.

NVIDIA hasn’t announced anything of this sort, everything posted online so far (such as supposed 6 or 12GB versions of the 780Ti) is rumors.
For many people, there would be little incentive left to buy their 3000$+ Quadro/Tesla GPUs if the 780Ti came out in these configurations. Gamers don’t need 12GB of VRAM, either.

The important part is to have this layer of abstraction exist at all, so not everybody who wants to develop an out-of-core solution has to completely roll their own. Coherent memory access is the key to good performance, with or without unified memory. In terms of raytracing, things could be done here with ray re-ordering and queuing.

I think there is such a card coming. Octane boss alludes to one a little while ago and again today. I think he might be in the know… :wink:
It wont compete with Tesla because it will be crippled for double precision like other Geforce as they cripple Geforce for pro apps in preference to Quadro. I doubt the intended market is gaming or scientific calcs, more like GPU rendering, movie effects etc

I am curious about how CUDA 6 will just for the 500 card series speed up GPU rendering. The switch to CUDA 5 was a nice speedbump already.

I posted a thread a while back asking about this but got no responses. Do you need to install anything to to access this cuda 5 speed upgrade or is it something that updates as your gpu updates?

Hi BrentNewton, nothing to do on your side.
Blender maintainers had change to Cuda 5.5 since 2.68, may we see change again to Cuda 6.0 for 2.70.
But Cuda 6.0 is Beta only atm…

Cheers, mib.

Oh thanks for the quick answer. You know I think I’ve seen a bit of a slowdown after this last update. Might just be imagining it though.

Wonder if the NVidia Force GT 120 card that my son clipped on the second slot will be enough to give Cuda any try (514 Mo…) … What do you think about?

Almux, the >GT 120 is to old, Blender support Cuda Capability 2.0 and higher.

Cheers, mib.

Interesting !

But do you think older cards will get the benefit (I own a 570, 2.5GB) ? Or will CUDA6 require a specific architecture only available on future cards ?

I thought I have seen a tech demo of clarisse/isotropics a bit ago, where a high resolution model of a Boeing was shown on a gtx 480 and the Ram (16gb) which had stored the mesh streamed its data to the gpuram.

but sadly I wasn´t able to find the clip anymore.

anyhow official Ram streaming possibility is a good thing and if it is available for current cards (from 5xx - 7xx and so on) the better.

is Mantle which is from AMD, going into the same direction?

Hey, the driver 331.40 support Cuda 6.01 and the OpenCL performance also rise 20% up.
Bad days for AMD.
The article is in german but interesting and a bit offtopic too:

Will test on my windows workstation.

Cheers, mib.

I dont think existing cards are going to be affected by “Unified Memory”
look on the nvidia roadmap or google for the nvidia news from last GPU Technology Conference