Samsung lays out a broader AI memory plan, from HBM5 to shared memory

Samsung is pitching a wider set of memory and storage technologies for AI infrastructure, not just faster high-bandwidth memory. The plan includes HBM5, shared memory for AI systems and SSD changes intended to handle the data generated by training and inference.

Watch Desk analysis

What happened

At the OCP Global Summit, Samsung presented three priorities: speed, memory expansion and data persistence. Its plans include HBM5 and customer-tailored HBM, plus zHBM, which stacks memory vertically on compute chips, and processing-in-memory for handling simple operations inside memory. HBM5 was shown as a concept mock-up and wafers, not announced as a product ready to buy.

For expansion, Samsung plans to use CXL memory modules and memory pooling so computing systems can share memory resources. It also demonstrated a CXL-based memory expansion solution with AMD. For storage, Samsung highlighted next-generation enterprise SSDs, QLC and flexible data placement, which it says can improve capacity and reduce SSD write operations.

These plans and demonstrations were described in Chosunbiz’s report on Samsung’s OCP presentation.

Why it matters

AI systems need more than fast processors: they also need room for working data, ways to move it between components and storage for large datasets. Samsung’s pitch is that memory capacity and data handling should be designed alongside bandwidth, rather than left to clean up after the accelerator has taken the spotlight.

The CXL demonstration with AMD is the most concrete sign of collaboration in the announcement. Sharing memory across systems could help address the capacity limits of individual servers, though the presentation does not provide deployment results or performance figures.

Our read

This is a useful map of where Samsung wants to compete across AI infrastructure, from memory beside compute to data storage. The interesting part is the system-level thinking; the less thrilling part is that a strategy presentation is still a promise of work to come. HBM5 was shown as a concept, and the practical value of the other technologies will depend on what customers can deploy and measure.

What to watch

• When Samsung gives HBM5 a clearer development and availability timeline.
• Whether its CXL memory solution moves from demonstration to customer deployments.
• What performance and endurance figures Samsung publishes for its next-generation SSDs.

Discussion spark: Should AI infrastructure makers prioritise faster memory, or focus first on making memory capacity easier to share across servers?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.