At the recent Arm Techcon event, Drew Henry, Arms VP/GM for Infrastructure, unveiled Arm's new brand for infrastructure called "Neoverse". The move by Arm represents a strong commitment to a new line of products aimed at the Infrastructure business: compute, networking and storage from the edge to the data center. To be clear, this launch encompasses IP of interest to the High Performance Computing (HPC) community.
Over the past 20 years, Arm has built a strong business licensing IP used at the edge in embedded, mobile and Internet of Things (IoT) devices. As process technology evolved, Arm evolved as the winning choice across a wide range of power and performance requirements. Today, Arm enjoys a strong position in the rapidly growing IoT/edge marketplace. In fact, Arm management paints a picture of a future with 1 Trillion edge devices.
This vision of the future, with 5G connectivity, involves a lot of data. So much so that the flow of data will reverses from today where outward bound video dominates, to tomorrow where inward bound machine data of many types (including video) overwhelms. This necessitates a new architecture (the fifth wave of computing) which sees inbound data growth of staggering proportions and an increasingly distributed network of computational resources (on top of ever more-capable edge devices) to handle the workload.
The new architecture is highly distributed, with rapidly growing edge processing needs in addition to FOG computing at points of convergence as well as data center cloud and of course High Performance Computing. Simple tasks, such as inference workloads will be pushed out to ever more capable edge devices. Where these devices need help, data will move to aggregation points (fog compute) or all the way back to the data center in the case of training workloads.
The HPC community has been working with parallel and distributed computing for over 25 years. Over that time, we have seen lots of change from the Cray vector systems and super-fast Gallium Arsenide technologies to the rapid rise of Beowulf computing clusters - the foundation of much of our HPC architectures today. Beowulf clusters leverage commodity class compute, a Linux OS, and standard HPC software that enables the distribution of applications and workloads across cores, sockets, and nodes in a local network. Arm is a newcomer to this space, but is already showing great promise in terms of compute density, performance per watt, efficient interconnects, and memory throughput. Expect some announcements from partners at SC18 in this regard.
A driving factor in this journey has been the cost of the world's fastest systems, which are the trend setters for the rest of the $30 Bln market called High Performance Computing. Today's systems are high end yes, but they are comprised of 10's or 100's of thousands of server-class sockets. A design that scales up as well as down to suit the particular need and environment. I would much rather be looking at the daunting task of putting 100K Arm sockets in a power-constrained environment than the current alternative. As Arm delivers on the Neoverse IP roadmap, our partners’ designs will bring industry leading compute per socket as well as per watt.
Another facet of maximizing peak compute in HPC has been the use of GPGPU’s. Being a general re-purposing of graphics cores, they aren’t the most efficient accelerators and they demand significant throughput in terms of memory and IO channels in order to operate at peak performance. As the calendar rolls forward, we expect that our partners will enable a variety of more modern, workload specific accelerators, all connected via a more efficient and coherent standard interconnect called CCIX. In some cases these accelerators will be large and reside in the data-center. In other cases the will sit out at the edge. In either case, they will provide efficiencies in performance yet to be seen in the world of green computing.
As Arm grows in the data center compute market, I'm pretty excited. First off, Arm is a serious player. With the volume of Arm-based products being created, the silicon industry has optimized manufacturing processes for Arm. Already you see Arm-based designs on 7nm technology and 5nm is not far off. Arm partners excel at designs that feature increasing performance in power-constrained environments and our ability to leverage this innovation in the HPC space will be a key driver to our success.
The announcement of the Neoverse brand and roadmap is clear indication that Arm is executing on the infrastructure side of its vision of 1 Trillion edge devices. In order to process the massive flow of data that those devices generate, the next generation architecture for the Internet, fog computing and the edge devices is evolving and HPC can not only play a major role, but will benefit quite handsomely as a result. You can expect to hear more from Arm in the data center and for HPC in the near future.
Join us at SC18 to hear more