Today, graphics processors (GPUs) and neural networks create the foundation for leading EDGE computing platforms, bringing AI and accelerated deep-learning performance as close as possible to industrial internet of things (IIoT) sensors in order to offer real-time processing, low latency, and high accuracy for applications such as smart cities and robotics. ECRIN Systems’ myOPALE-GPU building block (Figure 2) allows integration of the embedded Mobile PCI Express Module (MXM) GPU mezzanine directly to a Mini-SAS HD PCI board adapter, thus eliminating the fragile PCIe PC board connector, a legacy from the IT market that represents a huge point of failure in industrial PCs.
Embedded MXM GPU modules offer low power consumption and the thinnest available commercial off-the-shelf (COTS) solutions for high-performance parallel processing leveraging general-purpose GPUs (GPGPUs). For smart-city AI surveillance, fog computing with augmented reality, widening the field of vision for intelligent logistics, and Industry 4.0 in general, the NVIDIA GeForce MXM commercial-grade series platforms with Pascal architecture and RTX-2060 to RTX-2080 platforms with Turing architecture, offering ray tracing and deep-learning super-sampling technologies, are fine when your requirements do not include long life cycles (18 months only) and big-data processing.
For radar/sonar back-end computers, command/control human-machine interfaces (HMIs) in SWaP-constrained naval environments, test bench systems in aerospace, and medical ultrasound machines, you will prefer a solution offering a five-year life cycle with end-of-life product change notification (PCN) process support. Here, the rugged MXM embedded GPUs in the NVIDIA Quadro-grade series (Pascal P3000 to Pascal P5000 chip-down) operate from –40°C to +85°C and offer a new GPU Direct Remote DMA mode that supports the movement of big data for critical missions, with a bandwidth increase of up to 400% (from 3,500 Mbytes/second to 14,000 Mbytes/s) and a latency reduction of 500% (from 100 microseconds to 20 microseconds).
The myOPALE-GPU offering, which will formally launch in June at the Paris Air Show, has the same general characteristics as the myOPALE-CPU: 3U-inch box form factor, cold plate or passive heat sink depending on the environmental conditions, 12–48 VDC-only power supply, and PCIe Gen3 x16 lanes via two Mini-SAS HD connectors. On the front, there are four display ports if the user needs GPU output (not used for GPGPU massive computing). By connecting myOPALE-CPU with myOPALE-GPU, you can build a heterogeneous AI HPC modular computer composed of one Intel Xeon quad-core CPU module and two GPGPU modules via two PCIe Gen3 x16. Or you can use two CPUs with two GPGPU modules via four PCIe Gen3 x16 in a very short-depth 19-inch 350-mm chassis or short-depth 19-inch 450-mm chassis.