Trying Out the AMD Developer Cloud for Quickly Evaluating Instinct and ROCm

Last week alongside announcing the AMD Instinct MI350X/MI355X and the ROCm 7.0 software preview, AMD also introduced the AMD Developer Cloud as a new means for developers to easy try out Instinct accelerators with their own software and with the ROCm compute stack already setup. Having tried out prior AMD cloud compute environments, as soon as my email invite for the AMD Developer Cloud arrived I decided to give it a try.

AMD announced the AMD Developer Cloud as a new cloud offering for developers and open-source contributors. AMD awarded 25 hours of compute time to each of the attendees at their AI day last week in San Jose and will also be awarding developer credits to other open-source developers although the exact details and criteria for that haven't been made public yet. With having previously tested the AMD Accelerator Cloud and AMD Cloud Platform, I was curious to try out this new AMD Developer Cloud in wondering what is different about it, how easy is it to get going with Instinct + ROCM on this new cloud, etc. The standard pricing is $1.99 USD per hour for the single MI300X instance or $15.92 per hour for the 8 x MI300X as of writing.

The AMD Developer Cloud is accessible via devcloud.amd.com. Ultimately it redirects to DigitalOcean. The AMD Developer Cloud is relying on DigitalOcean for the cloud environment and ultimately deploying GPU Droplets.

As of writing, creating a GPU Droplet with the AMD Developer Cloud allows either creating a single GPU AMD Instinct MI300X or 8 x GPU Instinct MI300X instance. There wasn't any Instinct MI350 series access available in the AMD Developer Cloud when I was trying out my testing. The AMD Developer Cloud announcement at the AMD AI Day was rather brief but being alongside the MI350X/MI355X announcement I assumed there would be MI350 series availability to help in evaluating AMD's next-gen wares, but for now there is not and just the existing MI300X... Similarly, a bit surprising none of the MI325X hardware either. I was hoping there was going to be robust AMD hardware in the AMD Developer Cloud to evaluate the different AMD customer options. Hopefully this comes soon.

Also catching me by surprise with the AMD Developer Cloud was ROCm 6.4.0 being the only ROCm compute stack offered out-of-the-box. ROCm 6.4.0 released back in April and was succeeded by ROCm 6.4.1 in mid-May. ROCm 6.4.1 is the latest stable release while ROCm 7.0.0 formally rolled out into preview form last week though there was an initial ROCm 7.0 preview in late May for beginning to evaluate HIP C++ changes with this next major release. For the AMD Developer Cloud, I expected this to be shipping the very cutting-edge ROCm software.

I am told that the web UI for the AMD Developer Cloud was locked down before the final preview build was ready, but that soon the AMD Developer Cloud should be updated for the preview build. In any event it's worth pointing out that once your cloud instance (GPU Droplet) is deployed you can always install/upgrade whatever software at your choosing, but still for the AMD Developer Cloud hopefully they strive for delivering very timely preview/latest-release build options to enhance the initial developer experience on Instinct.

Besides being able to deploy with ROCm out-of-the-box, there are also quick start packages offered for JAX, Megatron, PyTorch, SGLang, and vLLM. It's nice having these options but again once you deploy your cloud instance you can ultimately install whatever packages you would like.

Ubuntu 24.04 LTS is the default operating system on the AMD Developer Cloud. It was quick and easy deploying a GPU Droplet with the AMD Developer Cloud. Again, powered by DigitalOcean and an easy-to-use web interface. Within minutes of entering the AMD Developer Cloud I was able to get an Ubuntu 24.04 instance running with ROCm 6.4.0 and eight AMD Instinct MI300X accelerators.

Catching me by surprise was these AMD Developer Cloud instances at least for now relying on Intel Xeon processors... In particular, the INTEL XEON PLATINUM 8568Y+ from the Emerald Rapids generation. For being the AMD Developer Cloud, I was thinking they were going to be powered by AMD EPYC especially with how the likes of the EPYC 9575F have been talked up and shown excellent potential for being AI CPU head nodes and the like.

So for those curious about the AMD Developer Cloud, that's the quick overview of this new developer offering from AMD. The AMD engineers I spoke with about the AMD Developer Cloud though don't recommend using it for any Instinct benchmarking / performance comparisons due to possible noisy neighbors, etc, so that's all for now without any performance benchmarks. At least the AMD Developer Cloud provides a quick and easy environment for those wanting to try out an Instinct accelerator with ROCm pre-configured in a matter of minutes. Moving forward hopefully they will be providing timely access to the Instinct MI350 series for evaluating AMD's next-gen wares and allowing developers to begin testing their changes against that forthcoming hardware. Similarly, hopefully they will be punctual in updating their OS images with the latest ROCm preview builds for providing the best possible out-of-the-box software experience