i us a laptop with 4gb ddr3LP and an arm processor for blender. No Cuda graphics either. Super slow but it works. The 472 GFLOPS on the Nano definitely would help.
but with 128 cuda cores…
its a cheap CPU so emulations like water and smoke should be slow but rendering will be ok i assume.
with blender I don’t know … but with machine learning, and cuda acceleration … this kind of projects should be fast …
I’m waiting for someone to come up with a working build of blender on these machines … I’m too curious to see what really happens
only with arm without opengl acceleration and cuda, it’s not even worth trying …
this machine, on the other hand, is a different story … that’s why I’m curious to see what happens
The ML Models are not tweaked, the device runs Tensorflow the same as all my other green GPU’s I am running local Deepracer training on 2 the same as I run them on a Titan, i see almost an imperceptible difference in the epoch training vs. the Titan (mind you the Titan is not at all stressed by the Deepracer training. It is basically an ARM computer with a smaller GPU which is useful. I am planning on setting one up with Blender this week
do you have this device?
are you able to compile and test blender 2.80 on this device ??
I would be very interested to hear which kind of models you have tried out. Also, which settings did you use for the training?
I have a few, at the moment I am using them for Machine Learning, I am going to setup blender this week hopefully but I am not good at compiling
I am running the Amazon Deepracer stack, I have got the Tensorflow poirtion running on the Jetson, I need to rebuild for the full stack but working through the ARM dependencies, ideally it will all be run on containers but the tensorflow portion can run the model training almost as fast as my main machines
Jetson Nano can run a wide variety of advanced networks, including the full native versions of popular ML frameworks like TensorFlow, PyTorch, Caffe/Caffe2, Keras, MXNet, and others.
I’m too curious about the results, good luck with blender this week …
keep us informed!
For inference it seems to be a reasonable solution. The numbers are for FP16 and a single batch for inference. Isn’t that too slow for training purposes? For what kind of tasks are you using it?
As long as we have a blender version that runs on arm Linux kernel that comes with Jetson, I would guess it should be doable. Jetson comes with all the gpu drivers, so you should get the hardware acceleration.
I was able to build blender on master for Jetson nano and it’s been working fine for me. GPU seems to be in use.
BTW, I needed to comment out a few lines that would enable SSE2, which is not available on ARM.
make a videorecording now and post it on youtube and here!
I’m too curious to see what happens!
the demo files you can use as a test are here
I have tried all the demo files found on blender.org. Blender on Jetson Nano was able to read all the files, but it struggled with some files. For example, the animation in the forest demo was flickering and not smooth. The frame rate was about two thirds of that I saw on my MacBook Pro (mid 2012 w/ 2.7GHz i7 and 16GB RAM). I monitored GPU status with an app ‘jtop’, and now I’m pretty sure GPU was in use.
Overall, mostly it works fine, but performance-wise, it was not impressive.
that some of these demo files could be obsolete by now (they were created when blender 2.80 was still very immature) maybe this is why there are problems … it is also possible, rather it is obvious that ARM needs optimization …
still good to know that it works …
(a 2 minute screen recording would be great)
cuda acceleration of cycles rendering works?
preferences-> system-> cuda.
you will have to do a test in a few days, when this function
DRW: Refactor to support draw call batching will be implemented, which will greatly reduce the reading between cpu and gpu …
surely you will have significant speeding up.
Could you save systeminfo of that nano (Help/Save System Info)? Does it recognize CUDA in GPU?
can you menage to run Cycles benchmark on those Cycles demos files and post results?
Does this help?
Looks like GPU was used only for display? I’m totally a newb to blender and latest GPU things so not sure. Thanks for the info btw, I’m going to update the repo and built it again next week.