Napatech, a leading provider of Altera-based programmable Smart Network Interface Cards (SmartNICs) and Infrastructure Processing Units (IPUs), used in cloud, enterprise, government, and telecom datacenter networks, disclosed its newest SmartNIC engineered for the challenging demands in artificial intelligence (AI) infrastructure. Napatech’s 400 Gbps N3070X AI SmartNIC is based on the latest technology from Altera, an Intel company.
AI creates new demands on datacenter servers and CPUs to support machine learning (ML), large language models (LLMs) and generative AI (GenAI). These workloads both utilize and create massive amounts of data that must be transported, stored, and secured. Even the most modern enterprise and cloud datacenter architectures have proven to be ill-suited to support the rapid changes from AI.
New AI-centric infrastructure requires a unique combination of capabilities:
- 200 and 400 Gigabit Ethernet, to transport massive amounts of data among AI clusters and nodes
- Ultra-low latency, for real-time, parallel, high-performance computing of AL/ML data
- Highly-deterministic delivery, with zero packet loss, and in-order delivery
- Programmability, to adapt to customer-specific environments and rapidly expanding set of use cases
- Openness, to conform to emerging multi-vendor standards across servers, NICs and switches
Network operators evaluating early network architectures for AI have been challenged by vendor lock-ins based on proprietary InfiniBand and RDMA over Converged Ethernet (RoCE) implementations. Napatech’s solutions conform to a vision for a standardized Ethernet infrastructure, utilizing Ultra Ethernet.
Also Read: SK hynix Develops PEB110 E1.S for Data Centers
Napatech and Altera have collaborated to solve these problems based on solutions architected by the world’s largest hyperscale cloud operators, and deliver them in production-grade, turnkey, products for the mass market. The N3070X is ideally suited for emerging use cases of AI/ML such as: AI/ML clustering, including FPGA and GPU interconnects; High performance computing (HPC) and Back-end storage. With the programmable AI SmartNIC, early adoption of Ultra Ethernet, a leading networking technology aimed at these AI-driven use cases, becomes possible.
Napatech’s N3070X AI SmartNIC is based on Altera’s powerful Agilex™ 7 FPGA technology. The 400 Gbps, x16 PCIe 5.0 SmartNIC delivers the raw speed, ultra-low latency, and determinism required by AI infrastructure. Compared to other proprietary solutions, it is hardware and software programmable to support unique customer environments, the exploding number of use cases, and the rapidly evolving technologies and standards. The N3070X supports several Agilex 7 FPGA options to deliver the optimal mix of price, performance, and functionality and is compatible with standard, high-volume servers.
The N3070X is designed for the future as it is hardware-ready to support an optional SoC board that features a powerful Intel® Xeon® D processor, to convert the SmartNIC into an IPU. In addition to the AI use cases, it also supports other popular workloads and acceleration for cybersecurity, fintech, storage, and many more.
“Traditional, basic or foundational Ethernet adapters and controllers are not designed for the complex networking environments that underpin the datacenter architectures used for artificial Intelligence workloads,” said Jarrod Siket, Chief Marketing Officer at Napatech. “AI infrastructure is an emerging use case fueling the increasing demand for Napatech programmable NICs including SmartNICs and IPUs.”
“In a future dominated by AI, Altera FPGAs sit at the intersection of growth and innovation in modern data centers, including their use as a coprocessor for CPUs, and as a networking engine for programmable NICs,” said Jeni Barovian, Head of Market Development at Altera, an Intel company. “Napatech‘s turnkey SmartNIC solutions based on Altera FPGAs help to enable new use cases driven by the explosion of Generative AI.”
Source: PRNewswire