Intel has announced its AI accelerator card Gaudi 3 is entering IBM Cloud and will be offered “as-a-service” for all customers.
Slated for open access starting early next year, this presents another option when it comes to helping enterprises scale AI more cost-effectively while giving those who had a soft spot for Team Blue some leniency when it comes to overall hardware and software support and compatibility.
As one of the earliest Cloud Service Providers (CSP) to add Intel Gaudi 3 to its lists of “as-a-service“ portfolio, the service will be made available for both hybrid and on-premise environments as well as IBM’s watsonx AI and data platform.
As generative AI continues to drive transformation, the demand for significant computing power underscores the importance of availability, performance, cost efficiency, energy efficiency, and security for enterprises. Through this partnership, IBM and Intel aim to reduce the total cost of ownership for scaling AI, while simultaneously enhancing performance.
Gaudi 3, when combined with 5th Gen Xeon, supports enterprise AI workloads both in the cloud and in data centers, offering customers greater visibility and control over their software stack, simplifying workload and application management. The joint effort between IBM Cloud and Gaudi 3 is focused on enabling customers to scale enterprise AI workloads more cost-effectively while maintaining a high priority on performance, security, and resiliency.
For generative AI inferencing workloads, IBM plans to incorporate Gaudi 3 support within its watsonx AI and data platform, providing watsonx clients with additional AI infrastructure resources for scaling their AI workloads across hybrid cloud environments, optimizing the price/performance ratio of model inferencing.
In short, the two industry giants aim to provide a full solution to help AI-driven clients with even the most stringent requirements obtain AI training and inference capabilities that are scalable and flexible.
Here’s a quick summary of Intel Gaudi 3:
- 1.8 PFLOPS of FP8 compute
- 64 TPC engines + 8 MME engines
- 128 GB of HBM2e memory
- 3.7 TB/s of HBM bandwidth
