In the Data Center space, FPGAs offer the low-latency offloading necessary to accelerate functions, such as Data Analytics, Artificial Intelligence, Smart Networking, Hyper-converged Storage and other functions. FPGAs support both in-line processing and look-aside to offload CPU workloads by reducing complex bottlenecks.
Software-defined storage (SDS) has become a rapidly growing industry trend. Remote storage systems are managed with a software-defined network responsible for replication and backup among other tasks to virtualize the storage network. This network can be improved via hardware with FPGAs to attain near native SSD latency, throughput, and IOPS despite being remote storage systems. With non-volatile memory express (NVMe) over PCIe*, SSD bandwidths reach nearly 3X or 4X better performance than over SAS/SATA respectively. This FPGA solution the Intel offers utilizes NVMe over RoCE, where Intel® FPGAs can act as both the host interface and the storage controller necessary in the SDS network, improving storage structures with native latencies for remote SSDs.
In RoCE implementations, storage area networks offload server network responsibilities to the FPGA relieving both CPU workloads and memory. With RoCE, server-to-server data storage transfer does not require the CPU and the process can be implemented with very low network latencies. In doing so, CPU memory buffers can be freed up for more pressing processing needs, improving the overall functionality of the servers.
For high-performance computing, FPGAs offer the low-latency offloading necessary to accelerate functions, such as design modeling, oil and gas search, nuclear power simulations, and other functions. FPGAs support in-line processing to offload CPU workloads by reducing complex bottlenecks. Additionally, FPGAs provide the ability for implementations of secure hash algorithm (SHA), de-dupe capabilities, erasure coding, and various forms of compression such as gzip. This in-line processing benefits system architecture dually by freeing up limited processor memory while lowering the computational load placed on the processor. By doing so, the FPGA can reduce power consumption and be in the optimized location for maximum performance in the data center.