FPGA Acceleration of Binary Weighted Neural Network Inference
White Paper FPGA Acceleration of Binary Weighted Neural Network Inference One of the features of YOLOv3 is multiple-object recognition in a single image. We used
RDMA
Low-Latency RoCE v2 at 100Gbps
The GROVF RDMA IP core and host drivers provide RDMA over Converged Ethernet (RoCE v2) system implementation and integration with standard Verbs API. The RDMA IP is delivered with a reference design that includes the IP subsystem itself, the 100G MAC IP subsystem, the DMA subsystem, host drivers, and example application on software. The system drivers are integrated with OFED standard Verbs API and are compatible with well-known RNIC cards and software. The IP core also provides a low-latency FPGA implementation of RoCE v2 at 100 Gbps throughput.
Compatible with RNIC and soft RoCE v2
100Gb/s throughput, 2µs latency
Configurable number of
RDMA queue pairs
Features
- Hardware operated RC, XRC, RD, UC, UD services
- Incoming and outgoing SEND, RDMA READ, RDMA WRITE
- Memory protection domains implemented in FPGA and ECN
- 3rd party MAC and DMA IPs
- Standard Verbs API on Host Machine
- Dynamic configuration using Verbs API
- Hardware retransmission and reordering
- Customizable IP
Enables RNIC use cases with FPGA-based SmartNIC
Block Diagram, Data Sheet and Product Details
Product Operation
The solution is a soft IP implementing RDMA over Converged Ethernet protocol. It consists of FPGA IP integrated with MAC and DMA, plus the host CPU drivers and is compatible with a variety of BittWare’s FPGA cards. The 200Gbps IP is compatible with BittWare’s IA-440i Agilex 7 I-Series FPGA card, and the 100Gbps IP is compatible with BittWare’s IA-840f and IA-420f Agilex 7 F-series cards and XUP-VV8 and XUP-P3R UltraScale+ FPGA cards. The solution complies with Channel Adapter and RoCE v2 requirements as stated in the IB specification. The diagram on page 1 shows a simplistic architectural overview of the system. The data plane and reliable communication is hardware offloaded and the implementation does not include CPU cores in the FPGA.
Detailed Feature List
- Fully compatible with known RNIC products and soft RoCE implementations (RoCE v2)
- 100 or 200 Gb/s throughput
- Configurable RDMA queue pairs
- 1023 or more
- 200Gbps IP: under 2.7 µs software to software latency (roundtrip) and under 1 µs hardware to hardware latency (roundtrip)
- 100Gbps IP: under 2.0 µs software to software latency (roundtrip) and under 300 ns hardware to hardware latency (roundtrip)
- Hardware retransmission management
- Memory protection domains implemented in FPGA
- Congestion control using ECN, PFC
- Can work with 3rd party MAC
- Dynamic configuration using Verbs API
- Standard Verbs API on host machine user / kernel space
- Hardware implemented Reliable Connection (RC), Extended Reliable Connection (XRC), Reliable Datagram (RD), Unreliable Connection (UC), and Unreliable Datagram (UD)
- Incoming and outgoing SEND, RDMA READ, RDMA WRITE
Reference Designs
The reference example consists of three parts:
- Encrypted FPGA IPs with reference design which implement RDMA protocol
- Software drivers which provide standard Verbs API support for the FPGA based RDMA adapter
- Example application build on top of the Verbs API demonstrating ping-pong test results: latency and bandwidth
Sample Implementation Results
| Device | LUTs | On-Chip Memory |
|---|---|---|
| UltraScale+ VU9P | 170K | 6Mb |
| Agilex 7 AGF014 | 170K | 6Mb |
Compatible FPGA Cards
Interested in Pricing or More Information?
Our technical sales team is ready to provide availability and configuration information, or answer your technical questions.
