The growing use of artificial intelligence (AI) in medical applications is leading to drastically higher performance demands for the embedded computing technology. Besides processing ever-increasing amounts of data — ideally in real time — medical devices must also handle local AI inference, record data from imaging procedures, and use data and cyber secure gateway functions to transmit these, for instance to PACS or the electronic patient record. At the same time, medical OEMs want to integrate their applications on a single hardware platform. To do so for a point of care environment, also requires functional safety. The new high-end COM-HPC computer-on-module standard provides the required technology basis for the development of the next generation of intelligent medical computers with integrated functional safety to manage critical medical device functions in the future.
AI is currently the major driver of medical advances in diagnostic imaging and treatment. It therefore comes as no surprise that market experts such as Transparency Market Research forecast a massive annual growth rate of 36.1 percent by 2031 to a market volume of $20.11 billion for AI in the medical imaging market.1 AI not only assists with medical diagnosis and the planning of follow-up examinations, but is also used for image acquisition, optimization, and processing. However, this means that the local systems must provide the necessary computing power to execute the AI inference — so intelligence for machine learning is not just required in the cloud or data center.
More Data — More Bandwidth
In addition, the demand for ever higher resolutions in diagnostic imaging and the resulting exponential increase in data volumes also drives up the performance needs. In endoscopy, for example, 8k resolutions are already fairly standard today even though they were only introduced in 2018. In this context, 8k means 37.8 megapixels at 8,192 × 4,320 pixels, which leads to four times higher data volumes than the 4k resolutions with 4,096 × 2,160 pixels that are common in other imaging procedures. When further considering that modern surgical procedures are to be performed by robots, possibly even using telemedicine, it is easy to understand that the systems must provide extremely fast real-time data transmission for controlling the robots. Processing an uncompressed video stream with 4k RGB resolution at 10-bit HDR color dynamics and 60 frames per second requires 14.83 Gbit/s; to do the same with 8k takes 59.33 Gbit/s.
More Tasks — Fewer Systems
But that is by no means the end of all requirements. To consolidate mixed-critical applications efficiently on a single platform, the individual workloads such as visualization, functional safety, and gateway should ideally be integrated in one system. Ultimately, this also reduces the hardware costs and the MTBF of the overall system. For security and certification reasons, however, the workloads must be decoupled through virtualization to ensure they do not influence each other. This, for instance, allows system components that are not relevant for certification, such as visualization or AI, to be updated without the entire system having to be recertified. It is also vital to ensure that the safety-critical control system is not affected by a crash of other system components. Besides many processor cores, this requires comprehensive virtualization support.
Modules Make Embedding AI Easier
Developers of the next generation of multifunctional medical computer systems must therefore not only integrate new core technologies; they also require an entirely new class of medical computing performance. To optimize the reliability and time-to-market of their new designs, they are also more dependent than ever on application-ready building blocks and design-in support for embedding the necessary intelligence and connectivity. Computer-on-modules — particularly those that comply with the PICMG’s new COM-HPC standard — are an important strategic lever for this.
Computer-on-modules are by far the most popular technology for embedded designs across a wide range of industries. Computer-on-modules are super components that integrate all important PC building blocks such as CPU, GPU, and RAM as well as a wide range of standard interfaces in a pre-validated, ready-to-use component. Any custom-specific parts are designed into a carrier board, which is a comparatively easy task, and the modules are then simply plugged onto that. As the modules are easy to swap, applications can be scaled and upgraded with current processor technology even years later, which significantly increases the ROI of expensive medical system developments.
COM-HPC is the latest standard and has been specifically designed to accelerate the development of the next generation of networked medical computers and edge servers. The computing power, bandwidth, and connectivity it offers is unmatched by any other existing computer-on-module standard. It supports all current highspeed interfaces up to PCI Express 5.0, Thunderbolt, and 25 Gbit Ethernet. Depending on the application, there are COM-HPC client modules with powerful, integrated graphics for traditional high-performance systems, or headless COM-HPC server modules for medical edge server applications that can be deployed, for example, as front-end or backend servers in CT or MRI systems.
COM-HPC Client for Standard Medical Designs
COM-HPC client modules are available in three sizes: 120 × 160 mm (Size C), 120 × 120 mm (Size B), and 120 × 95 mm (Size A). Modules based on the 12th generation of Intel Core processors (code-named Intel Alder Lake) are brand new. Size A and Size C COM-HPC modules provide several performance enhancements and functional improvements for medical applications. The most impressive among them is that developers can now use Intel’s innovative performance hybrid architecture. With up to 16 cores/24 threads, 12th generation Intel Core processors offer a quantum leap in multitasking and scalability.
Optimized for highest embedded client performance, the graphics of the LGA processor-based modules are now up to 94 percent more powerful. The inference performance for image classification has also almost tripled with up to 181 percent higher throughput.2 In addition, the modules offer massively more bandwidth for connecting discrete graphics processing units (GPUs) for maximum graphics and GPGPU-based AI performance. All other peripherals also benefit from doubled lane speeds, as they now feature ultra-fast PCIe 5.0 interface technology in addition to PCIe 4.0. They support up to four independent 4k displays.
COM-HPC Server for High-Performance Medical Server Designs
For applications that require data center level performance but no dedicated graphics output, COM-HPC server modules offer two solutions: Size E (160 × 200 mm) and Size D (160 x 160 mm). They are the ideal basis for use in medical backend systems that need to process massive video and image data in parallel, as well as in PACS servers and other radiological and hospital information systems. In this context, it is interesting to note that COM-HPC modules are not limited to classic x86 processors but can also integrate other computing accelerators such as FPGAs. This makes it easy to plug together high-performance visualization servers for MRI and CT systems.
COM-HPC-Based Multi-Module Designs: Reference Design for Machine-Learning AI Clustering
COM-HPC edge server designs are not limited to single-module concepts. The standard also explicitly supports multi-module carriers with heterogeneous COM-HPC module configurations that integrate, for example, FPGAs, GPGPU, or AI computing accelerators. A mix of COM-HPC server and COM-HPC client modules on one board is also possible. For example, congatec is currently working with the University of Bielefeld and Christmann IT on an edge server design that combines several COM-HPC modules on a single carrier board to handle extreme workloads or real-time requirements. The individual COM-HPC modules are connected via 10 Gbit Ethernet and PCI Express (host-2-host). This enables workloads such as fast and efficient AI clustering of high-dimensional data (self-organizing maps).
COM-HPC server Size E and Size D modules with BGA-mounted Intel Xeon D processors — which were developed under the code name Ice Lake D — are brand new. They impress not only with support for the extended temperature range of –40° to +85 °C. They also overcome many of the previous bottlenecks caused by edge server limitations and will therefore significantly accelerate the next generation of real-time workloads in harsh medical environments and extended temperature ranges. They offer up to 20 cores, up to 1 TB of memory on up to 8 DRAM sockets at 2933MT/s, up to 47 PCIe lanes per module in total, and 32 PCIe Gen 4 lanes with double throughput per lane, and up to 100 GbE connectivity and TCC/ TSN support. This provides ample bandwidth to transmit full resolution 8k HDR video streams over the existing Ethernet interfaces, eliminating the need for additional Ethernet controllers. Video storage and analytics servers also benefit from integrated Intel AVX-512, VNNI, and OpenVINO support for AI-based data analytics.
n-to-1 Thanks to Real-Time Virtualization
For deterministic real-time server consolidation, where diverse real-time and medical applications run independently on a single module, COM-HPC platforms support real-time virtual machines, for instance with the help of the RTS hypervisor from Real-Time Systems. All COM-HPC server-on-modules from congatec are prequalified for these services.
Co-creation for Medical OEMs
A highly convenient solution for OEMs developing IT/OT systems for the medical and healthcare sector are the co-creation services offered by congatec in collaboration with S.I.E. The systems engineering value of the cooperation between the two companies and their customers spans the entire supply chain, from computer-on-modules to series production of certified system platforms. The joint offering is aimed at medical device manufacturers and infrastructure providers who must meet patient safety as well as data and cybersecurity standards for the digitization of medical care.
The first examples of the computer-on-module offering and the power of congatec and S.I.E’s co-creation services are two new medical edge computing systems, called secunet medical connect Carna and Athene, which were developed and are currently being manufactured in collaboration with OEM customer secunet. These two systems are also examples of the high market maturity of the COM-HPC standard and the time-to-market advantages that can be achieved with computer-on-modules. Bearing in mind that the COM-HPC standard is barely two years old, it is particularly impressive that complete and certified medical servers are already available.
- “AI in Medical Imaging Market Trends, Growth, Insights by 2031,” Transparency Market Research.
- Intel Core i7-12800HE scores are estimated by Intel as of November 2021. Pre-silicon estimates are subject to +/- 7 percent error. Intel Core i7-11850HE scores are measured by Intel. Single-threaded performance measured with SPE-Crate2017_int_base (1-copy)IC19_0u4 (est). Multithreaded performance measured with SPE-Crate2017_int_base (n-copy)IC19_0u4 (est). Graphics performance measured with 3DMark Fire Strike graphics score. Configuration 1: Processor: Intel Core i7-12800HE, PL1=45W, (6C+8c) 14C, 20T, Turbo up to 4.6GHz. Graphics: Intel Iris Xe Graphics Architecture with up to 96 EUs. Memory: DDR5-4800 2x32GB. Storage: Samsung 970 Evo Plus (CPU attached). OS: Windows* 10 20H2, Windows Defender OFF, Virtual Based Security OFF. Configuration 2: Processor: Intel Core i7-11850HE (TGL-H), PL1=45W TDP, 8C16T, Turbo up to 4.7GHz. Graphics: Intel Xe Graphics Architecture with up to 32 EUs. Memory: DDR4-3200 2x32GB. Storage: Intel SSDSC2KW512GB (512 GB, SATA-III). Platform/ motherboard: Intel internal reference platform. OS: Windows 10 Pro 21H1, Windows Defender OFF, Virtual Based Security OFF. Bios: TGLSFWI1.R00.4151. A01.2104060640 (Release date: 04/06/2021). CPUz Microcode: 28h
This article was written by Farhad Sharifi, General Manager at congatec US (San Diego, CA). For more information, visit here . Contact: Farhad.