Both are based on a customise 16nm UltraScale+ fpga and therefore can be reconfigured to accommodate new algorithms and standard or shifts in workload needs.
The Alveo U200 has 892k look up tables (LUTs), 35Mbit internal SRAM and 31Tbyte per second internal SRAM bandwidth and can process 3100 images per second.
The U250 has 1341k LUTs, 54Mbyte internal SRAM and 38Tbyte per second internal SRAM and processes 4100 images per second to outperform GPUs and CPUs in cloud and on-premise data centre servers, said Manish Muthal, vice president, data centre, at Xilinx. He cited the example of the Aleveo U250 that can increase real-time inference throughput by a factor of 20 compared with high end CPUs and reduce latency by a factor of three, compared with GPUs, to accelerate machine learning and database searches.
Both cards are in production now and at the Developer Forum, visitors could see them in action.
One was gateway defined networking (GDN) company, Algo-Logic Systems, which demonstrated dashboards for smart cities and utilities. Its third generation Black Diamond Rackmount (BDR-3) acquires real-time data from analogue sensors to record motion (acceleration), power (voltage, currents), pressure (air, water), for example, by recording and streaming synchronised channels with high resolution over Gigabit Ethernet and uses fpga logic for signal processing.
Combined with the company’s Key Value Store (KVS) data fusion and scalable analytics, implemented on the Alveo U200, the integrated system can merge real-time analytics dashboards and integrate live data into immersive, 3D visualisations, for smart cities to monitor and control traffic systems, or for industry to monitor remote sites and operating systems.
The Alveo U200 accelerator cards with the Algo-Logic KVS can be installed in a standard PCIe slots, deployed in data centres, and/or accessed over the AlgoCentral cloud as a service.