Enfabrica has announced the availability of its 3.2 Tbps Accelerated Compute Fabric (ACF) SuperNIC silicon and pilot systems during Supercomputing 2024 (SC24). Designed to address the growing demands of AI workloads, the ACF SuperNIC delivers up to four times the bandwidth and resiliency compared to current GPU-attached NIC solutions. Initial quantities of the chips will ship in Q1 2025, positioning Enfabrica as a key player in AI infrastructure development.
The ACF SuperNIC is tailored for AI workloads such as training, inference, and retrieval-augmented generation (RAG) in large-scale GPU clusters. With 800-Gigabit Ethernet ports, support for 500K+ GPU clusters, and software-defined networking (SDN) capabilities, the solution enables efficient two-tier network designs. It also improves resiliency using proprietary Resilient Message Multipathing (RMM), eliminating network failures and job stalls without requiring software stack changes.
Enfabrica’s core Accelerated Compute Fabric (ACF) is built to optimize data movement and interconnectivity across heterogeneous compute environments, specifically tailored for AI workloads. The ACF technology powers its SuperNIC, which features high-radix connectivity with up to 160 PCIe lanes and 32 network ports per chip. This architecture enables efficient scaling of GPU clusters to over 500,000 GPUs with low latency and high bandwidth. The SuperNIC integrates advanced features such as Resilient Message Multipathing (RMM) for seamless failover, Collective Memory Zoning for zero-copy data transfers, and Software-Defined RDMA networking for enhanced flexibility and programmability. These innovations allow the SuperNIC to deliver superior network throughput, resiliency, and operational efficiency, catering to the demanding requirements of large-scale AI training, inference, and retrieval-augmented generation (RAG) tasks.
Enfabrica’s ecosystem partners, including OEMs and cloud providers, are expected to adopt the technology in 2025. The company’s innovative approach integrates high-radix networking, low-latency memory zoning, and enhanced RDMA functionality, ensuring high performance and flexibility for AI data center operators.
• 3.2 Tbps ACF SuperNIC silicon supports 800, 400, and 100 Gbps Ethernet interfaces.
• Designed to scale clusters to 500K+ GPUs with two-tier network efficiency.
• Features 160 PCIe lanes and 32 network ports per chip for flexible deployments.
• Utilizes Resilient Message Multipathing (RMM) for enhanced resiliency.
• Integrates software-defined RDMA networking for future-proof AI infrastructure.
“Today is a watershed moment for Enfabrica. We successfully closed a major Series C fundraise and our ACF SuperNIC silicon will be available for customer consumption and ramp in early 2025,” said Rochan Sankar, CEO of Enfabrica. “With a software and hardware co-design approach from day one, our purpose has been to build category-defining AI networking silicon that our customers love, to the delight of system architects and software engineers alike. These are the people responsible for designing, deploying and efficiently maintaining AI compute clusters at scale, and who will decide the future direction of AI infrastructure.”
- Enfabrica, headquartered in Mountain View, California, was founded by industry veterans Rochan Sankar and Shrinath Shivaram, both of whom bring extensive experience in networking and silicon design. Sankar previously held leadership roles at Broadcom and Marvell, while Shivaram has a background in building high-performance networking chips at Google and Marvell. Enfabrica has raised a total of $190 million in funding, with its most recent Series C round in 2024, led by notable investors such as Mayfield Fund, Sutter Hill Ventures, and BlackRock.
