Calling itself a leader in adaptive computing, Xilinx unveiled the Alveo U55C Data Center Accelerator Card and a new standardized API clustering solution for mass FPGA deployment at the SC21 supercomputer conference. The Alveo U55C accelerator is said to “deliver superior performance per unit of power in supercomputing and database workloads.” It scales easily with the Xilinx HPC clustering solution.
The novelty, specially designed for the tasks indicated above, is the most productive accelerator card in the entire history of Alveo, surpassing its predecessors both in computational density and in HBM2 memory, equal to 16 GB.
The advantages of the accelerator include high parallelism of data pipelines and optimized movement of data along pipelines. The card comes in a full height half length (FHHL) form factor. It occupies one slot, and its power consumption does not exceed 150 watts.
The above API-based clustering solution leverages RoCE v2 and bridging combined with 200 Gbps of bandwidth to form the backbone of Alveo’s network, which competes with InfiniBand in performance and latency. MPI integration allows supercomputer developers to scale the pipeline processing of data coming from the Xilinx Vitis unified software platform using Alveo. Leveraging existing open standards and frameworks, hundreds of Alveo cards can be bundled together, working with shared workloads and memory, regardless of server platforms and network infrastructure.
Software developers and data scientists can take advantage of Alveo and adaptive computing by leveraging the high-level programming tools in which Xilinx has invested heavily. Major AI frameworks such as Pytorch and Tensorflow are supported, as well as high-level programming languages such as C, C ++ and Python.
Alveo U55C card is already available. To familiarize yourself with its capabilities, you can use the FPGA-as-a-Service cloud services.