Touring the Intel AI Playground – Inside the Intel Developer Cloud

Intel Developer Cloud Infrastructure and Scaling

Aside from the CPU, GPU, and AI accelerator platforms, there is a lot more that goes into the Intel Developer Cloud. For example, companies that want to come in and train on Intel’s hardware need a place to store data. There were racks (three of them) with hard drives for cold storage.

There were racks of NVMe servers as well for hot shared storage.

Storage clusters, management clusters, Kubernetes clusters, and so forth are all connected via 100GbE, and Intel is also starting to connect more nodes with 400GbE as well.

A fun story we were told is that the Intel Developer Cloud started gaining momentum during the pandemic era when there was often a 12+ month lead time for high-end switches. As a result, the initial proof of concept version was using 1GbE. That is no longer the case, but it was a fun story.

It was pretty easy to tell which racks were running customer workloads and which were being setup quickly, as the cable management on the more mature racks was better. Something else we learned is that Intel is charging companies to come and use its infrastructure for training models. As such, it needs to run the Intel Developer Cloud like a cloud with security, monitoring, and high-availability operations.

Here is an example of some of the end-of-aisle network switches that were in a section being setup.

Managing access, given the fact it is hosting both paying AI customers as well as many users ranging from free accounts to paid developer access, is a challenge. We saw several racks dedicated to access like these Palo Alto Networks firewalls.

There was even an Intel-branded 1U security appliance.

The build-out is important because a lot is happening here.

AI Sweeping the Oregon Build-Out and More

One of the big changes to the Intel Developer Cloud in the AI rush has been power and cooling. Intel needed extra power brought to this section of the data center to bring more AI systems online. The future may be liquid cooling for these machines, but to get this running quickly, these systems are air-cooled, which was another challenge.

Next, however, was planned expansion that was about to be filled, but then changed. These look like the super cool racks from the reliability testing side, and Intel said it was about to build something here, but that the area is being retooled and repurposed to handle more AI servers.

This side of the Intel Developer Cloud cage needs to be used for AI servers because Intel is trying to keep all these AI machines close together to keep the runs to storage and other systems short. Plus, it has a lot of demand.

That is just the Oregon location.

A major challenge with AI is the data side. Regulations limit the portability of data, and latency is a challenge, so Intel is building much more than just this Oregon location. Indeed, it has a large Gaudi 2 super cluster already online running workloads for customers that is not at this location. We were told the plan is for this Oregon location to host thousands of AI accelerators soon. Beyond the Oregon campus, Intel also has projects for specific jurisdictions like Europe because data cannot be easily moved to an Oregon facility for its EU clients. Apparently, other locations are coming online as well.

Final Words

Between when we did this tour and when this article was published, Intel has put a lot of emphasis on the Developer Cloud. There is a ton here, and hopefully, this gives both those interested in hardware as a day job or just developers who want to know what is on the other end of those SSH sessions an idea of what the IDC is all about.

Personally, this was a fun project because we spent a lot of time asking to do this and getting approvals. I just wanted to say a quick thanks to the Intel team that made this possible.

For those out there who are looking for AI computing, or to try out new hardware you see reviewed on the STH main site and on our YouTube channel. Hopefully, this shows what Intel is bringing to the Developer Cloud and just some sense of the scale they are operating at. Many AI startups have demos from 1-64 accelerators. Being able to run customer production workloads on thousands of accelerators is something that the major cloud providers, NVIDIA, and only a few others aside from Intel can do.

We hope you liked this look at the Intel Developer Cloud in Oregon. We are going to have many more of these pieces coming soon!