As the military and defense sector continues to rapidly adopt modular, high-bandwidth platforms, process-laden systems capable of supporting a range of mission-critical applications are far more prevalent. These include areas from ISR and electronic warfare to radar, software-defined radio, geospatial processing and advanced video analytics.
The rapid adoption of AI into embedded systems now has GPUs playing a much larger role in overall system development. Instead of simply calculating how much GPU performance can fit into a system, designers are evaluating how the GPU works together with the processors, sensors, storage, networking, backplanes, software and thermal management.
AI-ready embedded systems used in these specialized applications rely on highly focused, trained models, using purpose-built, deterministic architectures where compute resources, data pipelines and models are optimized together. Far different from a consumer-focused generative AI perspective. (Figure 1)
Figure 1. AI-ready embedded platforms simplify integration, reduces time to deployment, and ensures mission-critical AI solutions at the tactical edge
System engineers need to evaluate how processing elements are connected and how easily they can be expanded, not just how an SBC pairs with a GPGPU.
SOSA profiles provide the framework for constructing AI-ready systems through interoperable plug-in cards, based off the VITA 65 OpenVPX standard. This existing ecosystem provides a relevant infrastructure to map GPU functionality across system development.
AI workloads vary dramatically, with the uses case dictating the system topology. A platform designed primarily to handle sensor data may need extensive I/O to support video, USB, Ethernet, expansion-plane capabilities and XMC expansion. Topology considerations of another performing complex inference may balance speed, lane width and protocol, along with the need to support multiple GPGPUs. A third may need to connect multiple accelerators, FPGAs or high-speed storage devices.
Depending on the application, designers can select either compute-intensive or I/O-intensive SBCs, along with GPU payloads and switches. The resulting topology must balance lane speed, lane width and communications protocols, while supporting the control, data and expansion planes required by the application.
As embedded systems become more data-centric, AI architecture continues to increase the amount of information pushed through a system. This can create other challenges to ensure the entire system can handle the data processing requirements of the GPU.
Surprisingly, it’s not the AI processing cores typically causing a bottleneck. As system development realigns to this influx of AI data, system designers need to focus on the potential limitations of memory capacity, data ingest and the buses moving information through the system.
In addition to the number of operations a GPU can perform, it’s also critical to also analyze:
Although not a new challenge, evaluating system performance needs against flexible operation takes on a different tone in the age of AI-ready embedded systems. At the heart are GPUs, which provide the massive parallel processing needed by edge AI applications, but are among the most power-hungry components typically used in embedded systems. Current-generation GPUs can easily consume in the 100W to 150W range, and even 200W or more - well above what a conduction-cooled 3U OpenVPX chassis in a deployed environment can be expected to cool. Thus, the integrator must push limits in two ways. First, they must optimize their chassis cooling design and integration strategy (e.g. deciding the order that cards are inserted into the backplane) to maximize the cooling for the GPU slots. Next, they must tune their application and reduce the GPU power consumption to the point where a) they are able to keep the GPU within thermal limits, and b) still accomplish their mission. This takes careful planning and experimentation up front in order to ensure a successful design. However, the benefits of using GPUs for the core processing of edge AI applications can be worth it both because of the processing performance that can be achieved, as well as the flexibility, enabling multi-function platforms and easy technology insertions.
Being availability and deployable are two distinct requirements. AI systems intended for military and rugged environments must work not only in the lab, but across temperature extremes, shock and vibration in real-world applications. And increasingly, developers want platforms that provide a realistic path from early prototyping to rugged deployment.
The AI-optimized CompacFrame development system from Elma addresses this engineering path, offering development and testing of high-performance embedded systems, which can then be quickly made field-ready. It includes six payload slots plus power, conduction cooling, internal fan cooling, up to 1,400 watts of onboard power, chassis management and maintenance-port aggregation. (Figure 2)
Figure 2. Elma’s AI-optimized CompacFrame enables rapid development of rugged GPU-focused applications next to the 3600S deployable rugged half ATR platform
The platform offers SBC options to meet different system needs. An I/O-intensive board can support application-specific expansion, substantial local storage and time-sensitive networking, while hybrid processor architectures can allow power to be shifted away from CPU performance and toward GPU-intensive AI workloads when necessary. Another typical application includes one GPU/SBC pair running an inference model and another supporting signal processing of RF data or driving a high-definition display.
This development-to-deployment model helps to reduce integration risk and accelerate technology insertion across military programs. Once the reference set has been developed and tested in the lab chassis, it is easier to move it into the deployable platform, like the Elma 3600S ATR shown in Figure 2.
It is not enough to merely put a powerful GPU into a rugged chassis. The platform must be ready to move large data sets at AI speeds, scale as workloads evolve, manage increasingly-demanding thermal and power budgets, and provide the security and interoperability required for deployment.
This shift in system development increasingly centers on AI systems optimization. And especially for the SOSA and rugged OpenVPX markets, AI capability is becoming a property of the entire standardized architecture, not a specification attached to one processor or GPU.
For a deep dive into the technical aspects of AI-ready system development, catch our on-demand webinar (link below): Getting Ahead of AI Processing Requirements with an AI Ready System Architecture for MOSA Development

在過去幾年中,模塊化開放射頻架構(MORA)已經發展,以通過基於開放標準的基礎架構來解決越來越複雜的無線電頻率(RF)系統的挑戰。由於多個產業合作夥伴共同合作開發合作框架,MORA 的互通性和模塊化已經實現,從而成功演示了多個製造商的技術共同合作。因此,我們詢問了我們的一些開放標準合作夥伴:現在已成功部署互通性演示,基於 Mora 的系統和嵌入式計算社區的下一步是什麼?

回顧後,我們現在可以看到開發平台的設計方式以及整合商客戶群如何使用它們的變化。這種轉變使執行可部署系統項目的開發階段變得更輕鬆、更便宜,並且比以往任何時候都更快地將解決方案交付戰士手中。開發硬體也可以在專案之間共用,也可以由後續專案繼承。這不僅節省實驗室預算,還可以節省訂購和接收新的開發項目所有新硬件的時間。