The new COM-HPC standard for high-performance Computer-on-Modules is about to be ratified by the PICMG. In this interview, Christian Eder, chairman of the PICMG COM-HPC technical subcommittee, reveals important details about the new COM-HPC modules, first samples of which are expected to hit the market during the first half of next year.
TimesTech: What are the key differences between COM-HPC and COM Express?
Christian Eder: Computer-on-Modules based on the new COM-HPC standard promise considerably higher transmission performance, many more high-speed interfaces, and significantly faster network connection, besides other benefits. This is down to a completely redesigned, more powerful new module to carrier board connector. While COM Express establishes this connection with 440 pins, the COM-HPC specification provides 800 pins. This doubles the maximum number of PCIe lanes from 32 for COM Express Type 7 to 64 for COM-HPC/Server.
And while COM Express supports a maximum of PCIe Gen 3.0 with 8 Gb/s per lane, a COM-HPC module achieves up to 32 Gb/s per lane via PCIe-5.0 – that’s four times the data rate of COM Express. COM-HPC modules will therefore be used in particularly performance-hungry applications, for instance to embed artificial intelligence with deep learning in embedded systems, or even to implement tactile Internet at the edge server level.
TimesTech: Speaking of edge servers: What performance enhancements can be expected from COM-HPC modules in terms of Ethernet connectivity?
Christian Eder: The enormous speed increase has an immense effect on the connectivity performance. Current COM Express modules (Type 7) at edge server level offer a maximum of 10 Gb Ethernet per signal pair. COM-HPC, on the other hand, specifies 25 Gb Ethernet (and more). With up to eight network connections, it then becomes possible to achieve transfer rates of 100 Gbit/s, and theoretically even 200 Gbit/s. Such rates are needed in the first instance for high-performance edge server solutions at the edge of telecom networks. Here, fast up, down and crosslinks in all directions must be established: i.e. north in the direction of the central cloud; east and west in the direction of neighboring edge fogs; and also south in the direction of industry 4.0 controls at process level. Another important aspect of high-performance computing is the amount of memory available. Depending on the selected module footprint, the COM-HPC standard offers up to eight DIMM banks on a single module to allow for substantial DRAM memory expansion, something which is often required for server applications.
TimesTech: You are talking about different module footprints. Does the new COM-HPC standard specify several module classes?
Christian Eder: Yes, just as there are currently Type 6 and Type 7 high-end specifications for COM Express, we have also planned two module classes for COM-HPC that address different application and performance requirements. In addition, there are two different form factors within these two module classes, similar to COM Express Basic and COM Express Compact. To be precise, we currently distinguish between server and client modules, in analogy to client/server computing. As mentioned earlier, COM-HPC/Server modules are tailored for use in edge server environments and require the largest possible memory capacity, a particularly powerful network connection, and the option to provide many cores for consolidating high workloads. These Server-on-Modules will host the mentioned eight DIMM sockets on a 200x160mm footprint, while the smaller 160x160mm server modules will integrate up to four DIMM sockets.
The COM-HPC/Client modules have a slightly more compact design, are also planned in two footprint variants with 120x120mm and 160x120mm, and are designed for use in high-end embedded computing applications. Unlike the server modules, they provide a maximum of 2x GbE interfaces (via NBASE-T) for Ethernet connection. In addition, COM-HPC/Client modules integrate video interfaces such as DDI and eDP/MIPI-DSI, which – in contrast to COM-HPC/Server modules – can be used to control up to four independent high-resolution displays.
While it is possible to use COM-HPC/Client modules comparatively power efficiently if necessary, the COM-HPC/Server modules designed for maximum connectivity and workload performance naturally come with a higher power draw. The client modules can be supplied with up to 200 watt, the server modules with up to 300 watt. For this purpose, each COM-HPC module specifies an impressive 28 power pins on the primary connector.
TimesTech: Does this mean the powerful COM-HPC modules will soon replace the COM Express standard?
Christian Eder: Certainly not. First of all, the reason for the development of a completely new COM module standard was of course the need to provide the hardware capacity required for the latest high-performance computing applications. When it comes to Server-on-Module applications at edge level, or machine learning with ultra-fast connections to graphics processors, even the fastest COM Express modules on the market today are reaching their limits. And it is exactly in these high-performance computing and top-level applications where the new, extremely powerful COM-HPC modules are being used, since COM Express modules simply cannot provide the performance required here.
This also explains why the COM-HPC standard will by no means replace the current COM Express standard. Instead, both specifications are likely to exist in parallel for many years and decades to come, depending on the field of application and its requirements. Remember – when the new COM Express standard was introduced 15 years ago, this did not lead to a rapid switch from ETX to COM Express either. Even today, there are new ETX modules on the market that continue to be used on carrier boards, some of which have been in operation for 15 years without a single relevant design change.
TimesTech: Who is responsible for the standardization of COM-HPC?
Christian Eder: COM-HPC is backed by the standardization organization PICMG (PCI Industrial Computer Manufacturers Group), which is also responsible for numerous other open standards in the embedded computing market. In it, more than 140 companies are working together on developing patent-free, open-standard specifications for telecommunications and industrial applications. The PICMG COM-HPC working group, founded on October 23, 2018, is a technical subcommittee of the PICMG and currently consists of 20 member companies. Since the group’s inception, the contents of the new COM-HPC standard are being planned, discussed and defined in weekly meetings. Next to the University of Bielefeld, PICMG COM-HPC members include Adlink, Advantech, Amphenol, AMI, congatec, Elma Electronic, Emerson Machine Automation Solutions, ept, Fastwel, GE Automation, HEITEC, Intel, Kontron, MEN, MSC Technologies, N.A.T., Samtec, SECO, TE Connectivity, Trenz Electronic and VersaLogic. Adlink, congatec and Kontron are also committee sponsors. I am the chairman of the COM-HPC committee and was previously involved as draft editor in the development of the current COM Express standard. Stefan Milnor from Kontron and Dylan Lang from Samtec support me in their respective functions as editor and secretary of the PICMG COM-HPC committee.
TimesTech: When is the standard expected to be fully ratified?
Christian Eder: The final pinout has just been released. From this moment onwards, there will be no more changes to the pinout. This means that the members of the COM-HPC working group will be able to start prototype designs promptly without having to wait for the final ratification of the standard, which is planned for Q1 2020. The first COM-HPC modules or previews can therefore also be expected to launch in Q1 2020.
