Empower your 200GbE FPGA farms with FPGA-based TCP/IP

Welcome to 200G FPGA-to-FPGA Connectivity: Scale Your AI Data Center with Reliable TCP/IP Solutions.

As data centers evolve, FPGA technology is emerging as a key component for intra-node communication, this empower modern FPGA farms with scalability and unmatched efficiency.

In these environments, speed is critical, and 200 Gigabit Ethernet provides the backbone for high-performance connectivity.

In this article, we’ll showcase a live demo of TOE200G Advance IP Core, demonstrating its ability to deliver the near line-rate performance in real-world FPGA-to-FPGA scenarios.

Data Center Intra-node with FPGA

In modern AI data centers, efficient inter-node communication is essential for peak performance. FPGA-based solutions provide seamless connectivity within computational farms, serving as the backbone of scalable Ethernet networks.

These solutions offer a fully integrated, all-in-one FPGA card, delivering unmatched ease of deployment and scalability. With reduced power consumption, high performance, and straightforward scaling, to transform data center infrastructure.

DG TOE200GADV IP Core

Design Gateway proudly presents the TOE200G Advance IP Core, a full TCP/IP offloading engine built entirely on FPGA. It’s designed for demanding workloads with Pure hardware logic without CPU, along with low power nature and Reprogramability.

Key features include:

• Supports both half-duplex and full-duplex transfers, achieving near line-rate performance.
• Handles up to 4 simultaneous TCP sessions, enabling flexibility for multiple applications.
• Configurable buffer sizes, up to 1 Megabyte, independently adjustable for each session.
• Provides standard user interfaces like Avalon Stream and AXI4 Stream for smooth integration.

Demo Setup

We’ll showcase the TOE200G Advance IP Core running on two FPGA boards: VHK158 and VPK120. These boards are connected via active optical cables, with each FPGA configured to run one TOE200G Advance instance. Let’s see the performance in action.

1 Session : Half Duplex

First, we demonstrate half-duplex transfer with one TCP session. The VHK158 board acts as the sender, and the VPK120 board as the receiver, transferring 512 Gigabytes of data.

Once the transfer is complete, the consoles show a throughput of 24,760 Megabytes per second, nearly achieving the 200 Gigabit Ethernet line-rate.

1 Session : Full Duplex

Next, we run full-duplex transfer with one TCP session. The setup remains the same, but now data flows simultaneously in both directions, transferring 1 Terabyte.

The result? The consoles confirm a throughput of 24,760 Megabytes per second, demonstrating full-duplex line-rate performance.

4 Session : Full Duplex

Finally, we showcase 4 simultaneous full-duplex sessions, representing 4 application streams simultaneously.

Once complete, the performance is totaling nearly 24,750 Megabytes per second. This demonstrates efficient bandwidth utilization across multiple data streams, ideal for scaling AI and cloud workloads.

Ready to take your AI data center to the next level?

Experience the power of scalable, high-performance connectivity with the TOE200G Advance IP Core. Visit us at design-gateway.com to learn more and transform your AI data center today.

?? Contact us