In the startup contest, we challenged technology startups, early-stage technology companies, and research institutions to use their expertise to solve real-world problems with innovative technology on Xilinx platforms. Participants used hardware platforms built around Xilinx’s 28nm, 20nm, and 16nm devices to demonstrate application acceleration with programable logic that showcases the significant performance or TCO advantages.
Over 55 startup companies worldwide participated in the contest. We received many fascinating entries, making it very challenging for our judges to narrow it down to the top three projects. We want to thank everyone that participated and say congratulations to our winners, who will be splitting a total cash prize pool of $180,000 (USD)!
First Place Winner
Project Solution: DeepField-SR: AI-based Super-Resolution, Accelerated by Xilinx Alveo Cards
Startup:BLUEDOT, is a South Korean venture company that provides fast and intelligent image processing and image compression technology using system semiconductors and artificial intelligence through cloud data centers.
DeepField-SR is a fixed functional hardware accelerator that leverages Xilinx Alveo cards and FPGA instances in the cloud or on-premise to offer the highest computational efficiency for video super-resolution. Based on the proprietary neural network trained with real-word video data from the internet and fusing spatio-temporal information in multiple frames, it produces superior high-resolution video quality.
DeepField-SR is deployable on both public cloud and on-premise with Alveo U200 and Alveo U50 accelerator cards. It’s designed in a scalable architecture and supports multiple Alveo cards to provide the flexibility to support various resolution upscale requests. Its runtime performance on a single Alveo U50 card is up to 14fps to upscale video up to 4K resolution. The API is integrated within an FFmpeg workflow, allowing a simple command to enable DeepField-SR acceleration and upscaling user input video.
3x Channel of Upscaling 360p30 to 1080p30/1440p30 in Real-Time
Second Place Winner
Project Solution: Yaddle: Molecular Dynamics Solution on Xilinx FPGA for Drug Discovery
Startup: Snowlake, headquartered in Shanghai China, is a leading technology company, providing FPGA-based specialized computing solutions to enable AI and HPC applications. The products have been widely applied to various scenarios including autonomous driving and high-performance computing solutions (seismic exploration and molecular dynamics simulation).
Snowlake Technology has developed Yaddle, a special-purpose solution for molecular dynamics calculations on Xilinx FPGA. Yaddle enables complete molecular dynamics computation on a single FPGA and efficiently outputs the results. With Xilinx Runtime Library (XRT), Yaddle provides compatibility support for commonly used molecular dynamics software and other analysis plugins via Yaddle APIs. On Xilinx Alveo cards, Yaddle (FPGA) achieves nearly a hundred times the performance of CPU and more than twice the performance of GPU. At the same time, Yaddle (FPGA) has only one percent of the TCO of CPU and one-seventh of the TCO of GPU.
Performance and TCO comparison for molecular dynamics on different platforms. For the simulation length(ns) for one million atoms in a day, with one server with eight Alveo U250 cards, Yaddle-MD can achieve 98x performance of Dual Xeon 9282 and 2.4x of DGX-A100.
Startup: Katoid Technology is a disruptive real-time analytics company founded in Barcelona, Spain.
Katoid's technology leverages the unique power of Xilinx FPGAs to process data from NVMe storage at wire speed. Horizontally scaled as an organic distributed system, Katoid allows web-scale real-time behavioral insights at computation speeds 100x faster than existing cloud solutions.
Katoid’s Interactive Analytics solution powered by Xilinx is available on Alveo cards, Zynq UltraScale+ MPSoC, and AWS F1 cloud instances. Its technology leverages the unparalleled computation power of Xilinx FPGAs, combined with the great bandwidth of NVMe SSD drives over PCIe, to offer 100x faster, 90 percent cheaper interactive big data analytics than competing solutions. It is possible to compute over gigantic datasets at interactive speeds and scale horizontally to power user exploration at terabytes per second.
Thanks to all our participants for making the first-ever Xilinx Adaptive Computing Challenge a huge success! Be on the lookout for the next Adaptive Computing Challenge opening in Fall 2021!