Chasing AI Inference, And Other Power Systems Stuff
February 3, 2020 Timothy Prickett Morgan
Without question, the Power AC922 is one of the best platforms for running HPC simulation and modeling workloads or machine learning training workloads, as is attested by the adoption of this platform for two pre-exascale systems at the U.S. Department of Energy dubbed “Summit” and “Sierra” by their respective Oak Ridge National Laboratory and Lawrence Livermore National Laboratory homes.
As you know from our past coverage of this platform, the Power AC922 gets the bulk of its compute capacity from the Nvidia Tesla GPU accelerators embedded in the system, which can have four or six of the “Volta” V100 GPU accelerators coupled to a pair of “Nimbus” Power9 processors over high bandwidth NVlink ports on both the CPUs and the GPUs. While this machine is good for HPC and AI training, it is not the right box for doing inference, and in fact, a machine equipped with the Tesla T4 accelerators from Nvidia is better suited for running the models created on the Power AC922. And that is precisely why IBM created the Power IC922 machine, which made its debut last week.
The IC922 was divulged in announcement letter 120-017, on January 28, and the machine will start shipping on February 7. The system can have six or eight of the Tesla T4 accelerators as well as a pair of Power9 processors with 12, 16, or 20 cores. The Tesla T4s are based on the same GPUs that are used in the newest GeForce RTX graphics cards, and the Tensor Core matrix math units in these chips are used to do inference or to do dynamic ray tracing. They are a lot less expensive than the Tesla V100s and are tuned for inference. This IC922 machine doesn’t run the PowerVM hypervisor or the IBM i operating system, so it can’t be used as a database engine (which is a pity), but given that it has 24 flash drives that will support the NVM-Express protocol, it will be a screaming data ingest engine and storage server as well as a screaming inference machine. The Tesla T4 accelerators don’t have NVLink ports on them and hook into the processor complex using PCI-Express protocols, and it is not clear if it can have all those flash drives and all of those T4 accelerators in them. I am guessing not, since there would not be enough PCI-Express lanes for that. But there are enough PCI-Express lanes to drive either the flash or the GPUs plus some fast networking thanks to IBM’s support of the PCI-Express 4.0 protocol on the Power9 chips.
If you want to know more about the feeds and speeds of the IC922, I detailed them over at The Next Platform, which you can read here. What IBM i shops need to consider is this: IBM is offering the IC922 at somewhere around 15 percent to 20 percent less cost than an equivalently configured base Xeon SP server with the same base CPUs, flash drives, and networking, and with flash and T4 pricings being roughly equal, IBM can compete for both storage server/database engine business and machine learning inference business. The Power9 machine can do somewhere around a quarter more transaction processing work, and that turns into 33 percent better price/performance. So if you have machine learning inference workloads, you really should give the IC922 a spin, just like you should consider the AC922 for HPC and machine learning training workloads. Anything that helps the Power Systems line fight again X86 servers is a good thing for the IBM i platform, after all.
In other news, IBM is rolling out new Ethernet switches for the Power Systems integrated solutions bundles, in this case a device that has 48 ports running at 25 Gb/sec and 8 ports running at 100 Gb/sec. These devices are detailed in announcement letter 120-003, and they cost $21,499.
Finally, IBM is withdrawing Red Hat Enterprise Linux 7.X support for its PurePower V7 appliance servers on May 1, you will see in announcement letter 920-017. No replacement is available, and frankly, we are surprised that the PurePower appliances are still being sold. The PurePower appliances were announced in June 2015 after IBM sold off the X86 server business to Lenovo, and IBM had big plans for these machines and even started supporting IBM i on them starting in October 2015. We have not heard much about them since then.