On the MUSTANG-T100, five Coral Edge TPUTM from Google® offer the possibility to execute up to five different AI tasks simultaneously. The performance of each TPU speaks for itself. 4 trillion computing operations per second at 2 watts power consumption per TPU, 2 TOPS per watt, in total according to Adam Riese up to 20 TOPs at 15 watts power consumption of the card. Instead of replacing existing systems with new ones, they can be sustainably upgraded with the Mustang-T100-T5. The five Coral Edge TPUs fit on a low profile PCI Express card with half height and half length. This makes it easy to integrate the Mustang-T100 into existing systems with a PCI Express x4 slot. In addition to x86 platforms, ARM platforms are also supported, as well as Windows operating systems and Linux. Upgrading existing systems thus enables a longer runtime and a reduction of the CO² footprint. The easy implementation of an AI model is guaranteed by Tensorflow Light. Systems with Mustang-TPU now offer parallel multitasking with which a separate AI model can be calculated per TPU. Shorter latency times can be achieved by model pipelining. This splits the AI model into segments to run the computation in parallel on each TPU, reducing the overall computation time. AutoML Vision Edge accelerates transfer learning, the execution of pre-trained models directly on the system, by using the TPUs. By transfer learning and optimizing the system with a small data set, accuracy is increased and training time is significantly reduced. Image classification, object recognition and image segmentation are ideal applications for the Mustang-T100.

ICP. Industrial Computer Products …by people who care!

Specifications

• Retrofittable AI accelerator card
• Low-profile PCI Express form factor
• 5 Google Coral Edge TPUsTM
• 2nd generation PCI-Express x4
• Multiple card support
• 15 watts power consumption

Applications

• Inference Systems
• AIoT Edge Computers
• Machine Learning Applications
• Object Recognition
• Face Recognition