Source36氪•Stalecollected in 2m
Samsung to Mass-Produce CXL 3.1 Memory Q4
Samsung CXL 3.1 mass prod Q4 – key for AI data center memory scaling.
30-Second TL;DR
What Changed
CXL 3.1 CMM-D memory module samples in Q3
Why It Matters
Advances data center memory disaggregation critical for scaling AI training and inference, potentially lowering costs for large-scale deployments.
What To Do Next
Test Samsung CXL 3.1 samples for memory pooling in your AI cluster prototypes.
Who should care:Enterprise & Security Teams
Key Points
- •CXL 3.1 CMM-D memory module samples in Q3
- •Mass production in Q4 post-certification
- •Targeted at servers and data centers
- •Supports next-generation memory standards
Deep Insight
AI-generated analysis for this event — not the original article.
Enhanced Key Takeaways
- •Samsung's CXL 3.1 implementation leverages the CXL Fabric Manager to enable multi-level switching, allowing for more complex memory pooling topologies than the point-to-point limitations of CXL 2.0.
- •The shift to CXL 3.1 is specifically designed to address the 'memory wall' in large-scale AI training by supporting fabric-attached memory, which allows for significantly higher memory capacity and bandwidth expansion per CPU socket.
- •Samsung is integrating its proprietary controller technology with high-bandwidth DRAM to reduce latency overhead, a critical bottleneck for CXL-based memory expansion in hyperscale data center environments.
Competitor Analysis
Primary Focus
- Samsung (CXL 3.1)
- Fabric-attached memory
- SK Hynix (CXL 2.0/3.0)
- Memory expansion/pooling
- Micron (CXL 2.0)
- Capacity expansion
Latency
- Samsung (CXL 3.1)
- Optimized for fabric switching
- SK Hynix (CXL 2.0/3.0)
- Standard CXL latency
- Micron (CXL 2.0)
- Standard CXL latency
Status
- Samsung (CXL 3.1)
- Q4 2026 Mass Production
- SK Hynix (CXL 2.0/3.0)
- Currently shipping 2.0
- Micron (CXL 2.0)
- Currently shipping 2.0
| Feature | Samsung (CXL 3.1) | SK Hynix (CXL 2.0/3.0) | Micron (CXL 2.0) |
|---|---|---|---|
| Primary Focus | Fabric-attached memory | Memory expansion/pooling | Capacity expansion |
| Latency | Optimized for fabric switching | Standard CXL latency | Standard CXL latency |
| Status | Q4 2026 Mass Production | Currently shipping 2.0 | Currently shipping 2.0 |
Technical Deep Dive
- CXL 3.1 Specification: Introduces enhanced fabric capabilities, including multi-level switching and improved peer-to-peer communication between devices.
- Memory Pooling: Enables dynamic allocation of memory resources across multiple hosts, reducing stranded memory in data centers.
- Latency Management: Utilizes advanced controller logic to minimize the performance penalty of the CXL protocol overhead compared to native DDR5 DIMMs.
- Interconnect Bandwidth: Supports PCIe 6.0 physical layer speeds (64 GT/s), doubling the bandwidth compared to CXL 2.0 (PCIe 5.0).
Future ImplicationsAI analysis grounded in cited sources
Data center TCO will decrease significantly by 2027.
Memory pooling allows for higher utilization rates of DRAM, reducing the need for over-provisioning memory in individual server nodes.
CXL 3.1 will become the standard for AI inference clusters.
The ability to share memory across multiple GPU/CPU nodes is essential for handling the massive parameter sizes of next-generation LLMs.
Timeline
2022-05
Samsung announces industry-first 512GB CXL DRAM module.
2023-05
Samsung develops 128GB CXL 2.0 DRAM module.
2024-03
Samsung showcases CXL 2.0 memory expansion solutions at MemCon.
2025-02
Samsung expands CXL ecosystem partnerships with major server OEMs.
- 2022-05Samsung announces industry-first 512GB CXL DRAM module.
- 2023-05Samsung develops 128GB CXL 2.0 DRAM module.
- 2024-03Samsung showcases CXL 2.0 memory expansion solutions at MemCon.
- 2025-02Samsung expands CXL ecosystem partnerships with major server OEMs.
Weekly AI Recap
Read this week's curated digest of top AI events →
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 36氪 ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.