Compute

The dedicated compute resources are model Dell R740XD servers with the following specifications.

CPU	2x Intel Xeon 12-core
RAM	192 GB
Storage	240 GB SSD
Network	2x 25 GB NICs
FPGA	Xilinx Alveo U200 Data Center Accelerator with 2x 100G ports
GPU	Nvidia Tesla V100 16GB VRAM (domains: bed, sb1, sb2) or Nvidia Tesla A100 40GB VRAM (domains: weeks)

Block Diagram

This block diagram shows the bus connections of the various processing devices hosted by these servers.

For best performance, you must take into account the data path, from source to sink, as well as processing and memory locality.

GPU

These servers each have an Nvidia GPU (V100 16GB or A100 40GB), for CUDA processing.

FPGA

These servers each have a Xilinx Alveo U200 FPGA, with two 100g qsfp28 ports. The FPGA can be configured either as an additional NIC, or as a standalone computing device.

In an advanced use case, you could use the FPGA to pre-process radio data, and then send over PCIe or the network to a GPU / CPU for further processing.

Network

In addition to the GPU and FPGA, the server has a two (2) port 25g ethernet NIC for data-plane traffic.

The 10G NIC is dedicated to control and provisioning traffic, and should not be relied upon for experimental repeatability.