IBM delivered the first of its POWER9 units to the Oak Ridge National Laboratories back in July 2017. This is a deployment that will take several months to complete. It also began deploying units to the Lawrence Livermore National Laboratory at the end of October 2017. These two announcements mean that customers and that includes IBM’s internal customers in SoftLayer and Watson are beginning to think about what Power9 means to them.
Cognitive – memory and processing hungry
One of the big workloads that IBM is focusing on for the Power9 is cognitive. Power9 and IBM Watson are a match made in heaven for IBM. Cognitive solutions require a lot of memory, processor power and performance. With Power9, IBM is looking to deliver all of this in a much smaller footprint than its rivals Intel and AMD.
IBM has announced two versions of the Power9 CPU that we know of. One has 12 cores with 8 threads while the other has 24 cores with 4 threads. That is a total of 96 threads per CPU and with the first servers likely to be dual processor machines, this then adds to a total of 192 threads per server.
But what does that mean? It’s a good question. Each thread is capable of executing a set of instructions. The more threads the more instructions that can be processed. However, most software today cannot take advantage of this large number of threads. That is because the software is not written or designed for parallel processing.
Most software works sequentially. This means that each process must wait for another to complete before it can run. In a world where there are limited cores and threads, it prevents the risk of software failing when a process times out. Parallel processing allows many processes to execute at the same time. However, this requires that the software is explicitly designed to take advantage of this approach.
When IBM Watson builds a cognitive model, it has to ingest and process extremely large amounts of data. Each piece of data needs to be processed. The more cores and threads that can be brought to bear the more data can be processed. With 24 or 48 cores per server delivering 192 threads, IBM is able to provide a dense platform capable of processing far more data in parallel than its competitors.
Enter the GPU
The evolution of the Graphical Processing Unit (GPU) from speeding up the processing of charts to enabling modern video games has been remarkable. They are capable of processing complex tasks far faster than traditional processors, including the Power9. GPUs also contain very large numbers of cores and are designed for parallel processing. Without it, they couldn’t render and deliver the video quality that modern computer games demand.
IBM has taken advantage of this. It has built its own accelerators into the Power9 processor that allow it to hand off processing to the GPU. Each Power9 server will support four and potentially more GPUs. As the GPUs are connected over a very high-speed connection direct to the CPU, the handover of data for processing is not limited by the speed of the system bus. The net result is that IBM can bring several hundred more cores, through GPUs, to bear on processing data.
But what about memory?
Data processing is not just processor limited. To deliver the data to the processor it has to be fetched from disk, loaded into memory and made available. For Watson, this is a critical issue. The data sets it can be working on are not measured in megabytes, or gigabytes. Most are in the order of terabytes and even petabytes.
The introduction of the Coherent Accelerated Processing Interface (CAPI) in Power8, and the much improved version inside Power9, IBM is able to treat flash storage as if it were directly attached to the processor. Data can be loaded into RAM and delivered to the processors at very high speed. As such, the processors do not end up waiting on the system bus to deliver data. It is able to create very large memory pools out of RAM and flash storage. This allows it to hold the data sets for building a cognitive solution in memory.
What does this actually mean for cognitive computing?
The first generation of Watson was limited by the CPUs and memory. Over the past four years, as Power has evolved so has Watson. One of the challenges for cognitive solutions is not just the processing of the data but the maze of intricate connections that are created between the data. It is those connections that create the cognitive solution.
To do this requires considerable processing and when something needs to be changed, there are two options. The first is to unload and reload the data, removing that which is faulty. The remaining data is then re-evaluated and recombined. The second approach is to do the changes in real-time. This is incredibly processing intensive and the current generation of systems can struggle to do this. As a result, the load-unload-reload approach is the norm.
For dynamic data sets where there are a lot of changes the load and reload phases can take increasing amounts of time. For IBM Watson for Cyber Security that time is offset by providing the service through the cloud with multiple instances. This allows IBM to constantly be rotating the instances taking one offline to update it while customers access the others.
Power9 takes IBM closer to the goal of not having to remove an instance from use while it updates. By combining multiple Power9 servers and running the cognitive solution over the top, customers will be able to consider a cognitive solution that is self-sustaining without downtime. This is a major breakthrough and one that will change how cognitive solutions are designed, delivered and used.
Directions for the customer base
Power9 has arrived and just in time before the close of 2017. That said, there is still an opportunity for IBM to do more work to tune Watson and Power9 ahead of the next release of AIX which is Q2/2018.
Customers should have the time to plan what they will want to do with Power9. Many may opt to use it in the cloud rather than on-premises. Cloud will certainly make it easier to scale up or down the amount of resource that they are using. For major data intensive tasks, Power9 is an attractive instance.
It also opens the door for more customers to consider their own cognitive solutions. Many are investing in chatbot technology to reduce their support load. Aligning this with Watson on Power9 means that the chatbot can take advantage of a cognitive environment which will give it a greater set of data when dealing with customers.