It would be nice if there was an AMI for the latest LTS version of Ubuntu with the CUDA drivers installed and working.
Last time I tried some CUDA work on AWS I spent the best part of a day trying to just get the drivers set up right for a basic test app to work on 14.04 - eventually I gave up and went back to 12.04 which, in 2015, didn't give me much confidence in CUDA as something I can expect decent support for.
I have this kicking around from some experiments a month or two ago:
sudo apt-get update
sudo apt-get install build-essential linux-image-extra-virtual
sudo reboot
echo options nouveau modeset=0 | sudo tee -a /etc/modprobe.d/nouveau-kms.conf
cat > /etc/modprobe.d/blacklist-nouveau.conf
blacklist nouveau
blacklist lbm-nouveau
options nouveau modeset=0
alias nouveau off
alias lbm-nouveau off
<Ctrl+D>
sudo update-initramfs -u
sudo reboot
sudo apt-get install linux-headers-$(uname -r)
wget http://uk.download.nvidia.com/XFree86/Linux-x86_64/346.35/NVIDIA-Linux-x86_64-346.35.run
chmod +x NVIDIA-Linux-x86_64-346.35.run
sudo ./NVIDIA-Linux-x86_64-346.35.run
nvidia-smi -q | less
This is obviously not production-ready, and is heavily cribbed from online (I couldn't quickly re-find where) but is good enough if you just want to play.
Canonical's packaging of the Nvidia drivers is risible.
It really is easier just to pull the driver and CUDA installers from the nv site and install manually.
Or use the Amazon Linux AMIs which just DTRT.
4 of the current Grid K520 GPUs in a single instance, at 4x the price (on-demand and spot) of the of the g2.2xlarge.
Shame they're not also rolling out a Maxwell hardware-backed instance type.
They're not even Grid K520s, they're K5000s. They've only got 4 gigs of RAM each, instead of the 5 gigs for teslas or 8 gigs for the K520s. This sounds a lot like they're trying to pull a little bit more value out of their current gpgpu rigs as they slide further and further into obsolescence.
Edit:
Looking again at the pricing and specs, it looks like they're just letting you have a full gpgpu box, instead of the multi-tenant box they previously offered where you share with three other people.
Well, the K520 is a single AIB with two discrete GPUs on it, each being a GK104 with 4GB (which seems to match the spec of the K5000). A g2.2x instance has a single GPU - ie half a K520 - assigned to it.
Your claim surprised me enough to stand an instance up and take a look for myself. Sure enough, they report having 4 K520 devices.
Gotcha. I was looking at the spec sheets of the various instances, and could only find that they are GK104s. I didn't think to look to see if there was a dual GK104 board, because video card makers tend to do that only for their high end boards. Definitely makes sense, though, for the density Amazon would want to run in.
Last time I tried some CUDA work on AWS I spent the best part of a day trying to just get the drivers set up right for a basic test app to work on 14.04 - eventually I gave up and went back to 12.04 which, in 2015, didn't give me much confidence in CUDA as something I can expect decent support for.