I use IBM Power Systems for online transactional processing, and in parts, also for online analytical processing.
We mainly use IBM Power Systems' IBMI, which is an operating system with an integrated database. For that system, the database is just another file system, which is ingenious because it saves you a full database administrator, or multiple database administrators, and that's a fairly cool thing.
They are reliable, their availability is very good, and their scalability is second to none. I can go from really small to very, very large, and still stay on the same platform, still go with a single server, rather than using multiple boxes and clustering them. To give you an idea, when considering Lotus Notes, that was an email workload, workflow software from IBM. When I ran that on a Windows X86 server, I could do something around 1,000 users per server. Beyond that, I needed additional servers. A couple years back, we introduced the same system for the US Department of Defense, and we had 275,000 users running on one and the same hardware, running on one machine.
There are multiple layers to the reliability aspect of IBM Power Systems. The hardware itself is very reliable due to the fact that I only have a very reduced set of hardware, so driver problems are pretty rare. I need to support two or three different ethernet drivers here and not thousands of ethernet drivers as with Intel. The hardware is very redundant, so high availability or disaster recovery becomes a breeze because normally, I just end up calculating one room, one local system, one remote system.
There are several features of IBM Power Systems that significantly improved processing efficiency. These machines come with built-in server virtualization, which has been there for the last 20 years. We started in 2005 with that. This allows me to put multiple workloads on the same machine, and those multiple workloads can communicate with each other very fast. In a normal network, I often encounter latencies because I need to travel through many cables, and I have switches in between that may not be available right away for forwarding my signal. With an IBMI, or with a Power System in general, I have a virtual network, meaning from one partition, I copy my data into the RAM, across to the next partition, and then put it through a network driver again. This is a latency-free network.
I am just starting with AI workloads and initiatives that are supported by IBM Power Systems. The reason for that is that IBM was slightly slow with supporting this. For commercial workloads, I wouldn't see anyone doing that before the end of this year because OpenShift Red Hat OpenShift for AI will not be available much earlier than that. I've been playing with that in a lab environment for about a year now, and it runs quite well.