Habana Labs, an Intel-owned AI company, has partnered with server maker Supermicro to provide high-performance, high-efficiency AI computing in the form of new training and inference servers that will power the upcoming NSF-funded Voyager supercomputer at the San Diego Supercomputer Center (SDSC). Established under a five-year, $5 million NSF award announced last year, Voyager is on-track to be available to users later this year.

Voyager will be a purpose-built AI cluster, containing both training and inference servers, based on Habana Lab’s Gaudi and Goya processors, respectively. The project is aimed at advancing AI research across a broad range of science and engineering areas.

The new Supermicro servers combine Habana AI chips and Intel Xeon CPUs in two configurations. The Supermicro X12 Gaudi AI Training System (SYS-420GH-TNGR) features eight Gaudi HL-205 mezzanine cards paired with two third-generation Xeon (Ice Lake) processors, while the Supermicro SuperServer 4029GP-T uses eight Goya HL-100 PCIe cards for AI inference, paired with two second-generation Intel Xeon (Cascade Lake) processors.

Voyager contains 42 training nodes, housed in seven racks, for a total of 336 Gaudi processors. In addition to the seven training racks, there is a networking rack that connects all the Gaudi processors through their RoCE (RDMA over Converged Ethernet) ports, as well as storage resources and two inference servers that are based on Goya.

Each Gaudi chip provides 32GB of HBM2 memory and implements 10 ports of 100 Gigabit Ethernet. Native RoCE provides intranode and internode communication.

Eitan Medina, Habana’s chief business officer, told HPCwire that Gaudi is the industry’s first, and he believes only, AI processor to integrate ten 100 Gigabit Ethernet ports of RoCE v2 on-chip, enabling flexible and cost-effective scaling.

“Because we designed the RoCE functionality from the ground up, we were able to squeeze 10 ports on each of the Gaudi chips natively,” he said. “In the Supermicro box, we connect the eight Gaudis directly over 100Gig links without any external components. Then from every Gaudi chip, we expose three out of the 10 ports to the box interface for a total of 24 ports of 100Gig that is available for node-to-node connection.

“This means you don’t need external NIC components and additional PCIe switches just to allow a Gaudi in one box talk to Gaudi in a different box. The interfaces come directly from the Gaudi processors and connect to the box interface.”

Medina contrasted the approach with how Nvidia’s DGX A100 systems are connected. “If you look at that [8-GPU] platform, which is very expensive, you need eight dedicated NICs in order to get eight ports of 200Gig. We have 24 ports here (without using NICs), so there’s a lot of advantages in integrating those RoCE links.”

The integration of training-plus-inference nodes on the Voyager system supports a unified workflow. Researchers will collect data, upload it to the Voyager storage servers and, according to the compute capacity allocated to them, use the Gaudi servers to train their model. Then, after the model is trained, they can start using the model by inferencing on new data.

Given SDSC’s research focus, the system is more training heavy, but will allow researchers to explore what can be done with high throughput inferencing with the option to add additional capacity as needed, Medina said.

Habana’s SynapseAI Software platform will facilitate training on the Gaudi processors. The platform supports popular machine learning frameworks, such as TensorFlow and PyTorch, and AI models for diverse applications, such as image classification, natural language processing and recommendation systems.

Founded in 2016, Habana Labs was acquired by Intel in 2019 for about $2 billion. Habana’s Goya inference processors have been available since 2019, but its Gaudi training processors are just coming to market. Last year, cloud giant AWS revealed it would be debuting Gaudi-based instances in 2021. This leaves Habana with a solid route to cloud and now to on-premises business opportunities.

“This is a major year for us,” said Medina. “The training solution, which is by far the more complex one, is getting to market through both the number one cloud provider as well as we are starting to see the on-prem installations.

“The announcement by AWS is a major validation, from our perspective, that will generate demand for Habana’s processors on both cloud and enterprise,” he said.

Medina said that on the ResNet-50 model, Gaudi delivers roughly 70 percent the performance of an A100 GPU from Nvidia, but that Gaudi offers superior price-performance. “The throughput per dollar at the system level is a worthy advantage,” he said.

Gaudi and Goya chips are both fabricated on TSMC’s 16nm process with 7nm products “coming soon,” according to the company.

Analyst Karl Freund, founder and principal analyst with Cambrian AI Research, told HPCwire that Habana’s progress is noteworthy.

“The design win with SDSC shows us that Habana Gaudi is the real deal,” said Freund. “One data point (AWS) is interesting, while two indicates a trend.”

Voyager is aimed at helping researchers tackle a broad range of AI science use cases. The first three years are structured as a “test bed” phase, where Voyager will be available to select researchers who will document and disseminate their learnings to support wider use. During the latter two-year phase, the resource will be more widely available via an NSF-approved allocation process.

“Machine learning techniques are rapidly becoming more prevalent across numerous science domains, from astrophysics to drug discovery and even the social sciences,” said SDSC Director Michael Norman, in a statement. “Voyager will significantly contribute to the high-performance computing community’s understanding of how such a system, built specifically for artificial intelligence, can be used to advance computational science and engineering research.”

“With innovative solutions optimized for deep learning operations and AI workloads, Habana accelerators are ideal choices to power Voyager’s forthcoming AI research,” said Amitava Majumdar, head of SDSC’s Data Enabled Scientific Computing division and principal investigator for the Voyager project. “We look forward to partnering with Habana, Intel and Supermicro to bring this uniquely efficient class of compute capabilities to the Voyager program, giving academic researchers access to one of the most capable AI-focused systems available today.”

Habana is one of several AI hardware vendors finding favor with supercomputing centers. Cerebras, which makes wafer-scale AI engines, has over the past two years deployed its CTS-1 systems at Livermore and Argonne National Labs, and been a partner on contracts with Pittsburgh Supercomputing Center and EPCC. SambaNova, another emerging AI silicon player, has had its SN10 RDU systems integrated into NNSA’s Corona supercomputing cluster at Lawrence Livermore National Lab. Taken altogether, today’s news is another shot in the arm for purpose-built AI silicon.