Yes, IBM’s Power11 systems use far more RAM than typical x86 servers—up to 8TB per socket—and no, they do not use CAMM modules. Instead, they rely on a proprietary form factor called OMI D-DIMM, built around the Open Memory Interface.
Here’s how IBM’s memory architecture stands out:
IBM’s OMI Memory Architecture
- OMI (Open Memory Interface) is a high-speed serial memory protocol developed by the OpenCAPI consortium (IBM, AMD, Micron, Samsung, etc.).
- It replaces traditional wide DDR buses with narrow, high-speed serial links, allowing up to 32 DDR5 ports per socket.
- This enables >1TB/sec bandwidth and up to 8TB of DRAM per socket, far exceeding what most x86 systems support.
Why OMI Beats Traditional DDR
- Bandwidth per die area: OMI delivers near-HBM bandwidth while using commodity DDR5, making it more scalable and cost-effective.
- Capacity per channel: Each OMI channel can support up to 256GB, compared to 64GB for DDR5 and 24GB for HBM2E stacks.
- Latency tradeoff: OMI adds ~6–8ns latency due to buffering, but IBM compensates with massive parallelism and bandwidth.
No CAMM Modules Used
- CAMM (Compression Attached Memory Module) is a JEDEC-standardized form factor aimed at thin laptops and some servers, offering better density and signal integrity.
- IBM does not use CAMM in Power11 or Power10 systems. Instead, they use OMI D-DIMM, a custom form factor optimized for their serial memory architecture.
Strategic Implications
- IBM’s memory subsystem is tuned for AI, HPC, and large-scale analytics, where capacity and bandwidth matter more than latency.
- Their approach is system-centric, not just CPU-centric—designing memory, interconnect, and accelerators (like Spyre) as a unified stack.
If you want to riff on how OMI compares to CXL.mem, or how IBM’s glueless 16P scaling affects memory topology, I’m ready to dive deeper.
Sources: Blocks and Files – OMI vs DDR vs HBM ServeTheHome – IBM Power11 at Hot Chips 2025 |
