FPGA Acceleration of Convolutional Neural Networks (CNNs)
White Paper FPGA Acceleration of Convolutional Neural Networks Overview Convolutional Neural Networks (CNNs) have been shown to be extremely effective at complex image recognition problems.
BittWare Partner IP
Arkville DPDK IP Core
DPDK/AXI Aware Data Mover IP Core
Computational packet processing applications such as software-defined networking (SDN), network functions virtualization (NFV), machine learning, video transcoding, image and speech recognition, CloudRAN, and Big Data analytics may employ both wirespeed gateware functions on an FPGA in conjunction with fast operations performed in software on one or more host processor cores.
Arkville DPDK IP core from Atomic Rules provides a high throughput line-rate agnostic conduit between FPGA hardware and GPP software. Using industry-standard AXI interfaces on the FPGA side and DPDK interfaces on the software API/ABI side, Arkville provides an exceptional “out-of-the-box” solution for both hardware and software teams. Because Arkville was designed with the specific goal of accelerating and empowering DPDK, the performance is significantly higher than one of a naïve DMA implementation on an FPGA.
Key Features
up to 220Gbps with PCIe Gen 4
Operate at any line rate up to 400GbE
DPDK and AXI standards
Features
- Offload server cycles to FPGA gates
- Bring your FPGA-based packet processing solutions to market quickly
- Future proof your GPP/FPGA application with the DPDK and AXI standards
- Line rate agnostic: Operates at any line rate, including 1/5/10/25/40/50/100/400 GbE
- Up to 220 Gbps and 500 Mpps with Gen4x16 interface
- 4 Physical Queue-Pairs (RX/TX) Standard; Up to 128 Physical Queue-Pairs
- Open-Source Arkville driver in DPDK 21.11
Arkville PCIe Gen4 x16 Throughput in Gbps
Video
View a demo of Arkville 17.05 performance plots.
Block Diagram, Data Sheet and Product Details
Product Operation
As shown in the block diagram, Arkville has both a hardware and software component. The hardware component is an IP core that resides in the FPGA, producing and consuming AXI streams of packets making ingress or egress. The software component is a DPDK PMD “net/ark”, the Arkville DPDK poll-mode driver. Arkville is a conduit between FPGA logic and Host user memory for bulk data movement or individual packets.
Together, an Arkville solution looks to software like a “vanilla” line rate agnostic FPGA-based NIC (without any specific MAC). DPDK applications do not need to change significantly in order to enjoy the advantages of FPGA hardware acceleration.
Detailed Feature List
Overall
- Ready-to-Go Solution to FPGA/GPP Packet Movement
- 4 Physical Queue-Pairs (RX/TX) Standard; Up to 128 Physical Queue-Pairs
- Single PCIe Physical Function (PF) supporting multiple ports
- Concurrent, Full-Duplex Upstream and Downstream Data Movement
GPP/Software Specific
- DPDK Arkville PMD in DPDK 21.11
- Tested extensively in with DPDK Test Suite (DTS)
- Unencumbered Application BAR (ABAR) for FPGA Application
FPGA/Hardware Specific
- AXI Streaming interfaces for packet movement
- Up to 256 Gbps, 500 Mpps burst traffic (Two 64 Byte wide, 250 MHz, AXI streams)
- Dedicated Application BAR (ABAR) AXI4-Master for the FPGA Application
- Integrated with Intel Quartus™ 21.3
Reference Examples
Atomic Rules provides Arkville example designs that may be used as a starting point for your own solutions. These include:
- Four-Port, Four-Queue 10 GbE example (Arkville + 4×10 GbE MAC)
- Single-Port, Single-Queue 100 GbE example (Arkville + 1×100 GbE MAC)
Sample Implementation Results
| Device | Speed | 6LUTs | FFs | M20k | Fmax |
|---|---|---|---|---|---|
| Intel Agilex F-Series | -2 | 81K | 220K | 250 | 500 |
Interested in Pricing or More Information?
Our technical sales team is ready to provide availability and configuration information, or answer your technical questions.
