Xilinx launched the most powerful accelerator in history, Alveo U55C, designed for HPC and big data workloads

     Xilinx, Inc., a leader in adaptive computing, announced at the Global Supercomputing Conference (SC21) the launch of the Alveo U55C Data Center accelerator card and a standards-based, API-driven clustering solution for large-scale FPGA deployment. The Alveo U55C accelerator delivers superior per-power performance for high performance computing (HPC) and database workloads while easily scaling through the Silings HPC cluster solution.

     Built specifically for HPC and big data workloads, the new Alveo U55C card is the most powerful Alveo accelerator card in Xilinx's history, capable of delivering the highest computational density and HBM capacity in the Alveo accelerator product family. Combined with Cyrines' new RoCE V2-based clustering solution, customers running massive computing workloads will benefit from enabling them to leverage their existing data center infrastructure and networks to implement powerful FPGA-based HPC clusters.

     "Scaling Alveo computing power against targeted HPC workloads is now simpler, more efficient and more powerful than ever before," said Salil Raje, Executive vice President and general Manager of Silings' Data center business group. At the architecture level, the FPGA accelerator represented by Alveo card can provide the highest performance for numerous computationally intensive workloads at the lowest cost. We introduced a standards-based approach that enables the creation of a Alveo HPC cluster leveraging a customer's existing infrastructure and network. With this approach, we are applying these key advantages to arbitrary data centers at scale. This is a significant leap forward in enabling Alveo and adaptive computing to be more widely used in the data center."

     The Alveo U55C card incorporates many of the key features required by current HPC workloads. It delivers higher data pipeline parallelism, superior memory management, optimized data migration across the pipeline, and the highest per-unit power performance in the Alveo product family. Alveo U55C card adopts a single-slot, full-height, half-length (FHHL) shape, and the maximum power consumption is as low as 150W. The Alveo U55C delivers superior computing density compared to its predecessor, the dual-slot Alveo U280 card, and also doubles the HBM2 capacity to 16GB. The U55C provides higher computational power with a smaller profile size, helping to create dense clusters based on the Alveo accelerator. It is designed for high-density streaming data, high I/O math, and large computational problems that require performance scaling, such as big data analytics and AI applications.

     By leveraging RoCE v2 and data center bridging technology, combined with 200 Gbps bandwidth, this API-driven clustering solution enables Alveo network to compete with InfiniBand network in terms of performance and latency without vendor locking. MPI (Information Transfer Interface) integration enables HPC developers to extend the Alveo data pipeline with the Silings Vitis™ unified software platform. Leveraging existing open standards and frameworks, performance can now scale across hundreds of Alveo cards, regardless of server platform and network infrastructure, while sharing workloads and storage.

     Leveraging high-level programming for applications and clusters, software developers and data scientists can leverage the Vitis platform to unlock the strengths of Alveo and adaptive computing. Silings has invested heavily in the Vitis development platform and tooling process, which aims to make adaptive computing easier for software developers and data scientists without hardware expertise. The Vitis platform supports mainstream AI frameworks such as Pytorch and Tensorflow, as well as high-level programming languages such as C, C++, and Python, enabling developers to build domain solutions with specific apis and libraries, or using the Silings software development suite, This can easily accelerate critical HPC workloads within existing data centers.