Search papers, labs, and topics across Lattice.
SCENIC is a new open-source datacenter SmartNIC that aims to bridge the gap between high-bandwidth commercial SmartNICs and customizable research SmartNICs. It achieves this by treating the NIC datapath as a stream computation substrate, integrating on-datapath Stream Compute Units (SCUs) with embedded ARM cores and direct access to GPUs/SSDs. SCENIC demonstrates performance matching commercial platforms while enabling versatile use cases like offloaded collective communication and network-to-GPU data partitioning.
SCENIC delivers the best of both worlds: the high bandwidth and software integration of commercial SmartNICs, plus the customization and data processing offload capabilities of research prototypes.
Although modern, AI-centric datacenters heavily rely on SmartNICs, existing devices impose a hard trade-off. Commercial SmartNICs provide high bandwidth and easy software integration, but offer limited support for customization and data processing offload. In contrast, research SmartNICs often suffer from low bandwidth, limited functionality, and poor software compatibility -- to the point that many are not actual NICs in a technical sense. This gap can be closed by treating the NIC datapath as a first-class stream computation substrate with shared hardware/software abstractions for a tight co-design of infrastructure and applications. To demonstrate this, we introduce SCENIC, an open-source datacenter SmartNIC. SCENIC implements a 200G network datapath over offloaded TCP/IP and RDMA stacks, as well as a fallback path for processing arbitrary network traffic. On top of the network logic, SCENIC combines on-datapath Stream Compute Units (SCUs) for data processing and embedded ARM cores for flexible control path manipulation with direct access to GPUs and SSDs. SCENIC is fully integrated with the OS, exposing native Linux network and RDMA verb interfaces, making the programmable datapath transparent to existing applications while enabling control of, e.g., user-defined offloads and programmable congestion control. SCENIC's performance matches commercial platforms, and we show its versatility through several use cases such as offloaded collective communication and network-to-GPU hash-based data partitioning.
ETH