GPU module delivers 17.7 TFLOPS, 68 dense/136 sparse Tensor TOPS
Stephen LawElectronics Embedded Systems Semiconductors computing embedded GPU military Processor rugged systems
CURTIS-WRIGHT Defense Solutions division VPX3-4936 3U OpenVPX GP GPU processor module features the combination of NVIDIA’s Ampere GPU and a configurable Gen4 PCIe Switch. The NVIDIA Ampere architecture offers a significant boost in performance and efficiency over the previous NVIDIA Turing generation, including more flexible concurrent execution of floating point and integer streams. The Ampere device’s third generation Tensor cores deliver up to 4x acceleration of AI/ML algorithms and its next generation RT cores and CUDA core architecture provide 2x performance compared to the previous generation. While delivering close to 18 TFLOPS FP32 peak performance and 68 dense/136 sparse Tensor TOPS, the NVIDIA Ampere also significantly improves power efficiency, yielding an extremely efficient 154 GFLOPS per Watt. Module’s PCIe Gen4 architecture also doubles the host interface bandwidth, eliminating data throughput bottlenecks. Pin-compatible with firm’s Turing architecture VPX3-4935, product enables system designers to boost math-intensive processing algorithms without increasing size, weight, and power (SWaP). The board’s PCIe architecture also supports Non-Transparent Bridging (NTB) and daisy chain options for system flexibility. Rugged module is designed in compliance with the U.S. Army’s C5ISR/EW Modular Open Suite of Standards (CMOSS) and is aligned with the SOSA Technical Standard to support compute-intensive ISR and EW systems.