At the Flash Memory Summit (FMS), Xilinx is showcasing a new member of its adaptable data center accelerator family, the Alveo U50.
The U50 card is the first low profile adaptable accelerator with PCIe Gen 4 support, designed to supercharge a broad range of critical compute, network and storage workloads, all on one reconfigurable platform.
The Alveo U50 provides a programmable low profile and low-power accelerator platform built for scale-out architectures and domain-specific acceleration of any server deployment, on-premise, in the cloud and at the edge. Xilinx claims that Alveo U50 delivers between 10-20x improvements in throughput, latency and power efficiency. For accelerated networking and storage workloads, the U50 card helps developers identify and eliminate latency and data movement bottlenecks by moving compute closer to the data.
Powered by the Xilinx UltraScale+ architecture, the Alveo U50 card is the first in the Alveo portfolio to be packaged in a half-height, half-length form factor and low 75-Watt power envelope. The card features high-bandwidth memory (HBM2), 100 gigabit per second (100 Gbps) networking connectivity, and support for the PCIe Gen 4 and CCIX interconnects. The 8GB of HBM2 delivers over 400 Gbps data transfer speeds and the QSFP ports provide up to 100 Gbps network connectivity. The high-speed networking I/O also supports applications like NVMe-oF solutions (NVM Express over Fabrics), disaggregated computational storage and specialized financial services applications.
Xilinx claims the Alveo U50 can provide significant latency and throughput benefits over existing CPU and GPU solutions in a wide variety of workloads, including machine learning inference, video transcoding, data analytics, computational storage, and electronic trading and financial risk modeling.
According to data provided by Xilinx, the Alveo U50 delivered 25X lower latency with 10X higher throughput than an NVIDIA Tesla T4 in a speech translation, deep learning inference workload. Running the TPC-H Query database benchmark, the Alveo U50 offered 4X higher throughput per hour than an Intel Xeon Platinum 8260 processor (at significantly lower cost), and it delivered much lower latency and higher efficiency than a couple of other Xeon and Xeon + Tesla configurations in electronic trading and financial modeling workloads as well. A computational storage acceleration test also showed 20X more compression/decompression throughput, faster big data analytics, and >30% reduced cost per node when compared to CPU-only solutions.
The Alveo U50 is sampling now with OEM system qualifications in process. General availability is slated for fall 2019.