This news blog provides news about the e-IRG and related e-Infrastructure topics.

Back

ARM is not just focusing on the Flops but on the science behind the workloads in HPC

During the Workshops Day at ISC'17 in Frankfurt, Germany, we were able to interview Eric Van Hensbergen from ARM. ARM is an IP design company. They develop IP for micro-processors. The traditional business was in embedded systems and mobile phones. Almost every mobile phone has ARM IP in it. The company is based in Cambridge, England but ARM has design centres all over the world. Eric Van Hensbergen is working at the design centre in Texas. There is one in Sophia, France, and there is the centre in Cambridge. ARM recently started to expand the markets that it targets with its designs, to include servers, and as part of that target, HPC.

Eric Van Hensbergen's background is in HPC. He joined ARM over five years ago to help lead the high-performance computing research effort. They started in high-performance computing in Europe with the Mont-Blanc project which continues to be a successful way to expand ARM's knowledge about HPC and improve the software ecosystem for HPC for ARM. The company is also involved in government programmes in the United States to target their exascale programme. A couple of years ago, the Japanese announced that they are also pursuing HPC with ARM designs.

ARM is not producing the processors and the hardware themselves, they just provide the technology. ARM as a company doesn't manufacture anything. The company provides designs and design services, which are licensed to silicon providers that make computer chips and those are then integrated into the systems. As a researcher and as a systems company, ARM spends a lot of time interacting directly with end users, application developers, and so forth, so that the ARM people make sure they understand what they are designing these systems for and designing this IP for. In this way, they can participate in the co-design and create much more efficient systems.

The Mont-Blanc project team has announced that they would make one product commercially available. The Mont-Blanc project has several partners and one of them is Atos/Bull. They produced early prototypes that were research-only prototypes based on ARM but now, at the ISC Conference, they have announced a product based on an ARM processor.  

The Mont-Blanc project is one of the European projects which are developing new technology for HPC in the direction of exascale and ETP4HPC is organizing its workshop at ISC17. How is ARM positioned into that group, we wanted to know.

Eric Van Hensbergen replied that ARM has been part of the Steering Board for ETP4HPC since before he joined the company. ARM is a big believer in development of technology and technology work force. ARM has long been a beneficiary of the European Union's involvement in improving technology. ARM's interests are very within ETP4HPC but ARM is absolutely here to help Europe develop strong European-sourced products, whether they be silicon designs, system designs or software. In all of these capacities, ARM is interested in the development of European industry and high tech.

The next phase within this European project would be something like system demonstrators which show the capability of a future exascale system. Is ARM also intended to be involved in that, we asked.

Through the many years that Eric Van Hensbergen has been involved with ARM HPC as a researcher, the ARM team spent a lot of time doing workload analysis and architectural development targeting high-performance computing with ARM. The company has developed a rich amount of knowledge about how ARM thinks its partners could develop high-performance computing to target different applications. ARM really wants to be involved in the process to help its partners be successful. Ultimately, some of ARM's revenue comes from licensing but most of it comes from royalties. ARM is only successful when its partners are successful. It is in ARM's best interest that the partners are successful, not just for producing prototypes but for producing products. ARM is very interested in the success of its partners.

ARM works together with Fujitsu to expand the IP for HPC with a vector-type of possibility. Are there any other HPC extensions, we inquired.

Eric Van Hensbergen answered that there are scalable vector extensions that ARM develops in coordination, not just with Fujitsu but also with other industrial partners, and both national laboratories in the US as well as national institutes in Europe. It is probably the most significant thing and actually a change to the ARM instruction set, specifically targeted at high-performance computing. That is the only thing that has been publicly announced. There are many things in ARM's research pipeline where the company continues to work on things to improve efficiency of data movement, synchronization, scalability, and the ability to couple on-chip interconnects with various forms of off-chip interconnects like communication networks and so forth.

One thing that has been announced in that area is the following. There were a couple of standards bodies that were involved in targeting specifically that for high-performance computing and data analytics. One of those technologies is something called C6 which is an interface standard, meant to address some of the deficiencies of PCI when one is talking to accelerators, FPGAs, or communication networks. There is another complementary standard called Gen-Z. ARM is also part of that consortium which again is a high-performance interconnect for talking to a variety of peripherals, memory and so forth. Both of these technologies are very interesting in high-performance computing because they will enable ARM customers to couple ARM processors with accelerators, large-scale systems, and interesting memory hierarchies which are going to be critical for achieving things like exascale.

In Europe there are now a lot of things happening in the HPC world, such as the EuroHPC announcement, the announcements from the German Ministry at ISC17. It seems like there is a re-focus on HPC and a determination to get somewhere in the timeframe 2020-2023. We wanted to know whether ARM is happy with that.

One of the things that makes Eric Van Hensbergen happy about this is that when he became interested and involved in HPC, earlier in his career, it was about the science that HPC enables. There is something rewarding about providing tools that allow us to expand our knowledge and improve the quality of life. HPC always seemed a path for him, as an engineer and as a scientist, that really had a tremendous capability in these areas. Europe has been one of the few places in the world that has stayed focused on the science. This is what HPC is about. It is not about some benchmark number. That is something that ARM has always been very focused on in its activities in HPC. It has been focused on being driven by the actual workloads and the science behind those workloads as opposed to just looking for a benchmark.

Eric Van Hensbergen was excited that Europe has started to accelerate the momentum in this area because they have held true to this goal of science, improving the quality of life, and improving industry and all of these things, as opposed to just focusing on Flops.