Was not able to get an OpenGL enabled GUI to run on this platform so i wasn't able to test the features i needed to

The drivers that are shipped with this image are not particularly up-to-date, even though the image apparently has been updated recently.

Most of the nvidia stuff has been installed into /opt/nvidia, but other stuff in /usr/lib64. Which is not a problem per se, just inconstent and a minior inconvenience.

Some things are just broken for development purposes, like the CUDA pkg-config file, which returns this build bot directory as location for the includes:

$ pkg-config --cflags cuda-6.5
-I/builddir/build/BUILDROOT/cuda-6.5.14-1.12.amzn1.x86_64/opt/nvidia/cuda/include

Lots of little things like that.

Having said that, it's not totally unusable.

I couldn't access my instance.
I am using AWS for two year and have already spent 20KUSD, but I couldn't find any help regarding this kind of instances.
I tried with java web service shell and with ssh shell.
I tried also to access my instance volume by attaching a copy to an ubuntu instance but the procedure I was usually using wasn't working.

Rgds
Victor

This Amazon Linux is ugly. Extremely old soft (for example boost is of 1.53 version, which is nearly 2 years old, the current one is 1.57). Many of useful libraries cannot be installed using yum (opencv, gflags-devel, glog-devel, lmdb-devel and other requirements for Caffe for example). NVidia driver is rather old (but not so old as boost). I'm disappointed. A lot of time is needed to just install all these libraries manually. Ubuntu offers much better user experience.

Out of date drivers for this machine, not worth updating as does not work! Installed latest version and got errors.

Hallo,

We wait 4 days for GPU Instances LIMIT Increase and we decide to build our own server with GPU for testing. So we do not have a chance to test it.

I've managed to get my software up and running on this AMI, but the lack of documentation and general inaccessibility of Nvidia's SDKs made it a lot more painful than it needed to be. The actual development environment for CUDA applications was bare bones, without the "GPU Computing SDK" that for some reason is separate from the "CUDA Tools" system, and is currently impossible to download from Nvidia's site.

I ended up having to spend a lot of time ferreting out different versions of the different Nvidia downloads until I could find a set that worked well together, when I'd hoped that an official AMI would come pre-provisioned with the foundations I needed, or at least some docs on how to set them up.

Using oclHashcat (only GPU power) about 27000 Hash/s, using Elcomsoft EWSA (GPU+CPU) about 28000-29000 Hash/s.
Required up to date version of EWSA to use Grid GPU, older doesn't recognize accelerator!

price wise
it is a fair deal as long as the job runs just for only a few hours.

Performance, however, is similar to a GForce 680 in terms of
GF/S so in case you have one, there is no point renting this instance.
Also a more recent GForce Titan or Radeon R9 has more computational power. Furthermore, it's just one card and not two like the previous version running two Tesla cards.

OpenCL performance is rather disappointing which seems to be inherently rooted to the Kepler Architecture. If you're developing CUDA only, this is not an issue but for porting from another GPU/APU
is not a good idea at all.

I thing Amazon should also consider AMD Radeon instance so that people have choice and can compare what kind of GPU fits their Job best.

That ~$0.65 an hour translates to $15 a day, ~$110 a week or over 400 bucks a month. For regular usage, you get already a much more powerful GPU for 400 bucks...

Do the math for yourself and ask yourself how much you are willing to pay for your GPU task.

Would I rent it again? Properly not.
Would I rent another instance with more GPU power? Maybe yes.