Gaudi and Gaudi2-based Servers Deliver Flexibility with Industry-standard Interoperability
State-of-the-art deep learning applications require multiple processors in systems with multiple inter-card links running at high speed and with ample bandwidth. To provide flexible, high-speed interconnect across multiple ASIC solutions the OCP OAM (OCP Accelerator Module) specification defines the industry-endorsed form factor to enable interoperability across multiple, compliant ASIC solutions. Habana Gaudi2 mezzanine cards comply with the OCP-OAM 1.1 specification, giving customers the flexibility to interchange processor equipment, resulting in greater control over the long-term for infrastructure components and costs. Gaudi2-based servers are being promoted at this week’s 2022 OCP Global Summit in San Jose, CA.
Three new Gaudi2 server solutions feature the Habana Gaudi2 deep learning processor, which demonstrated leading deep learning time-to-train in the June 2022 MLPerf benchmark, outperforming the Nvidia A100 by roughly two times on ResNet-50 computer vision and BERT natural language processing models. Each of the servers contains eight Gaudi2 mezzanine cards, built to the OCP-OAM 1.1 specification. The Gaudi2 mezzanine card supports the OCP-OAM interconnect and integrates 24 100 Gigabit Ethernet ports with 21 ports dedicated to an all-to-all configuration, connecting to the other seven cards within the servers. Three ports are dedicated to scale out, delivering a total 2.4 Terabits of Ethernet connectivity per Gaudi2 server.
Habana HLS-Gaudi2 Deep Learning Server from WiWynn
This week at the OCP Summit, WiWynn is showcasing the new SV600G2, designed with OAM-based Habana® Gaudi®2 to provide extraordinary AI training throughput performance and ensuring customers with freedom from vendor lock-in. For more information, see the WyWynn press announcement and whitepaper.
Supermicro Gaudi2 AI Training Server (SYS-820GH-TNR2)
The Supermicro Gaudi2 server, featuring OCP-OAM compliant cards and interconnects, will be available this quarter to customers requiring high-performance deep learning solutions with the flexibility of the industry standard OAM for on-premises systems. The SYS-820GH-TNR2 is “on tour” in the Supermicro Rack-Scale Green IT Solutions Road show. See the video. For more information on the Supermicro Gaudi2 AI Training Server, see the website.
Inspur Gaudi2-based Deep Learning Server NF5688M7
Inspur is also offering a new AI server based on open computing specifications that integrates Habana Gaudi2 deep learning processors. The Gaudi2 HL-225H mezzanine card is utilized in Inspur’s next-generation OAM platform to accelerate deep learning development on open architecture and supports a variety of computer vision and natural language processing workloads. See today’s announcement from Inspur.
For more information about Gaudi2-based solutions, see the Habana website.
To learn how you can obtain Gaudi2 servers for your own deployments, contact us or one of our partners featured at today’s OCP Global Summit.