IBM Gives A Peek Of The Future At POWERUp 2019
May 20, 2019 Timothy Prickett Morgan
It would not be a COMMON, or even a POWERUp, conference without some glimpse into the future by IBM to give customers of its Power Systems line a sense of what lies ahead near the horizon. By doing so, Big Blue can provide comfort to customers that it is working on future technologies and services without revealing its hand too much to competitors.
Steve Sibley, vice president of offerings for the Cognitive Systems division, which is the part of IBM that makes and sells Power Systems iron, participated in the opening session of the POWERUp 2019 conference in Anaheim on Sunday morning, and revealed a bit about upcoming products as part of a general talk about how IBM was pushing customers to do the second wave of digital transformation within their enterprises while at the same time embracing new technologies such as machine learning on Power Systems machines outrigged with GPU accelerators from Nvidia.
The talk of upcoming technology was set against a backdrop of a revitalized Power Systems business, which as we have previously reported has had six straight quarters of revenue growth. “And not only that,” said Sibley, “but as a testament to the modernization and the innovation that customers are doing with their businesses, we have actually seen six straight quarters of growth on the IBM i platform itself. This is the platform that we keep reinvigorating and delivering new capability around, and that customers are able to take and innovate with to deliver new business services.”
Here is how we might explain it: The System/38 was revolutionary, but few knew it and it was far too expensive compared to other systems at the time. The AS/400 was revolutionary to those who didn’t know the System/38 and was considerably more affordable compared to other platforms, and since that time, the AS/400e, iSeries, System i, and IBM i platforms have been steadily evolutionary from that base, and to such an extent that over the course of three decades, the machines we have today do not look anything like the machines from prior decades, but they can, if need be, still run the databases and applications created on those older instantiations of the platform. When you think about it that way, this truly is a remarkable product.
There are a number of different things that are driving the Power Systems business in general and the IBM i business in particular, and IBM is also focused on trying to bring more and more cognitive computing to midrange customers, which includes various kinds of predictive analytics and data analytics as well as new fangled statistical and deep learning machine learning routines that can sift through enormous amounts of data and create algorithms automagically. (This includes various Watson software and services.) Those newer deep learning and machine learning workloads are largely theoretical at this moment, although there are some customers just starting to embed machine learning techniques in their enterprise applications as enhanced recommendation engines and such. But this will be an increasing part of the workload at all types and sizes of enterprises. There just is no question about it. The only question is how machine learning in its many forms will be implemented at these companies, not if or even when. Companies might do machine learning training in the cloud and embed inference engines in their production servers, or use features of future Power processors to run inference.
We think the latter will happen in the IBM i base, and that is why we think that the vector math engines in the future Power10 processors will be able to support a wide variety of mixed precision integer and floating point math, not just 32-bit single precision and 64-bit double precision floating point. The lower bit levels allow for fuzzier data to be used in machine learning training, and thus for more data (of a lower resolution) to be pushed through the system, and time and again with machine learning algorithms, more data yields better results than higher precision data. Sibley did not talk about this in his keynote address, and certainly did not say that Power10 would support INT4, INT8, bfloat16, and FP16 data formats as well as FP32 and FP64, but we think that this will happen.
Sibley did confirm that the Power10 processor was still on track for delivery in 2021, keeping more or less to the three to four year cadence that IBM has between Power chip generations. Very quickly, Sibley talked about an announcement for what he called “hybrid memory” that would be coming either later this year or early next year, which we think is an enhancement to the Power9 chip – probably to be called the Power9’ and that is a “prime” sign to designate that it is a change in I/O not a “plus” sign as in the Power5+, Power6+, or Power7+ that was used to indicate a change in chip manufacturing process and tweaks to the microarchitecture. (The Power8 chip was primed up with NVLink interconnects to GPUs that supported better memory atomics and coherency across the CPU and GPU memories in a hybrid Power-Tesla compute complex.) We are not precisely sure what this hybrid memory is that Sibley is talking about, but it could involve using a mix of DRAM and other non-volatile storage in the memory complex as well as a new kind of buffered memory that speaks over the OpenCAPI buses and foreshadows the way that the industry will very likely deploy DDR5 main memory. We will obviously be following whatever this is.
The other tidbit that Sibley let slip is that the IBM Cloud just last week put slices of Power S922 and Power E880 servers into beta running IBM i 7.2 and 7.3 and AIX 7.2 and 7.3. IBM had pre-announced its support of IBM i and AIX on its public cloud back in March, and it is taking a bit more time than I expected to get this into general availability. But we are convinced that in the long run, Big Blue will have lots of IBM i slices available in its cloud.
Perhaps equally significantly, Steve Will, the IBM i chief architect, said last year at POWERUp that search engine giant, public cloud provider, and OpenPower Foundation founder Google was going to be firing up IBM i slices on its Google Cloud Platform public cloud. And Sibley confirmed in his keynote that Google was still indeed planning to offer IBM i and AIX slices on GCP at sometime in the future. It is not clear when this will happen, but Google has said it is running Power9 gear in its datacenters to support its internal applications, and it is not much of a stretch to put PowerVM on its “Zaius” homegrown servers so they could run IBM i and AIX. Google may want an open source variant of PowerVM for its internal use and to hook it into its GCP systems, which are based on the KVM hypervisor, or perhaps IBM and Google have worked together to port IBM i and AIX to the KVM hypervisor.
The latter would be very interesting indeed, if it turns out this is the way IBM i comes to the Google cloud.