Maximize Performance with IBM’s Cloud-Native AI Supercomputer
A lot has happened in artificial intelligence this past year, from new technologies being well-known to models with tens of billions of parameters being used for practical purposes. The sophisticated AI capabilities were brought to IBM clients across a wide variety of industries with the launch of Watsonx, the data and AI platform for enterprises. This was made possible by utilizing several inventions that originated from our IBM Research community.
An increasing number of tasks across the AI lifecycle necessitate the development of systems with appropriate computational capacities. For this and other reasons, IBM built the AI supercomputer Vela last year and put it in the IBM Cloud. By utilizing Vela, we can streamline the entire AI workflow on the IBM Cloud. This includes data pre-processing, model training and tuning, deployment, and even incubation of new products.
Read: 4 Common Myths Related To Women In The Workplace
What is Vela?
Vela was built to be versatile and extensible, so it can train current large-scale generative AI models and handle future demands as they come along. Its architecture was also made to be easily deployed and maintained from any location on Earth. Vela has been used for AI training and prototyping by IBM AI practitioners over the past year. One such AI technology is watsonx.ai, IBM’s next-generation AI studio. It would have been impossible to swiftly launch a global platform like watsonx.ai without Vela’s cloud-first design.
New: 10 AI ML In Personal Healthcare Trends To Look Out For In 2024
Features of Vela The AI Supercomputer
At this time, Vela is exclusively available to members of the IBM Research community. IBM has described the system as its new “go-to environment” for AI researchers within the firm. Though IBM did say as much, they also hinted that Vela is more of a trial run for a bigger rollout.
Details first. In contrast to IBM’s Power10 chips, which were announced in 2021, each Vela node has two Intel Xeon “Cascade Lake” CPUs, eight Nvidia A100 GPUs (each with 80GB of RAM), 1.5 terabytes of RAM, and four 3.2 terabytes of NVMe storage. The system is announced by IBM on its blog. According to the company, the nodes are linked through “multiple 100G network interfaces.” To ensure strong cross-rack bandwidth and component failure isolation, each node is connected to a separate top-of-rack switch, which is connected to four separate spine switches. One of IBM Cloud’s features is that Vela is “natively integrated” with their VPC environment.
Read: Data monetization With IBM For Your Financial benefits
USP
- Boosting capability
We wanted to make sure that adding more GPUs to Vela didn’t take up too much room or resources, even if that was the original plan all along. Specifically, we sought a method to quadruple the density of the server racks, which increased capacity by a factor of two without necessitating more room or networking hardware.
- The enhancement of diagnosis and operations
Improving the system’s efficiency was another goal of the Vela team. The complexity of AI servers causes them to fail more often than many more conventional cloud systems. Also, their failure modes are often surprising and difficult to identify. Also, a training job that uses hundreds or thousands of GPUs can have its performance affected when just one or two of the nodes in the network go down. So, to maintain the environment productive, automation that uncovers and detects these kinds of problems and generates alerts as fast as feasible is crucial.
- Boosting Vela’s speed
To train and deploy this specific wave of AI, there are unique interdependencies with the underlying infrastructure. The increasing size of data sets being used to train models necessitates the utilization of additional GPUs per operation to move quickly. To avoid GPU-to-GPU communication becoming a bottleneck to workload progress, there needs to be a corresponding increase in network performance as the number of GPUs computing in parallel grows. We successfully scaled training individual workloads to thousands of GPUs per job after deploying a major upgrade to the Vela network this year. We implemented GDR (GPU-direct RDMA) and RoCE (RDMA over Converged Ethernet) as the primary enabling technologies on Vela.
Read: Top 10 Benefits Of AI In The Real Estate Industry
[To share your insights with us as part of editorial or sponsored content, please write to sghosh@martechseries.com]
Comments are closed.