Marvell has announced two CXL 3.2 memory expanders for AI infrastructure, adding six memory channels and support for sharing memory across as many as four hosts. The company says the new Structera X controllers can expand capacity for larger models and longer context windows, with sampling expected in Q1 2027.
Marvell Technology Watch analysis
What happened
Marvell’s Structera X 3406 and 3506 each add six DDR4 or DDR5 memory channels, 50% more than the previous generation. The company says the controllers support up to 9TB of DDR4 or 6TB of DDR5 per controller, with up to 256GB/s of host bandwidth over PCIe 6.0.
The announcement also describes inline LZ4 compression, which Marvell says can provide two to three times more effective memory capacity, depending on workload. The company says its LLM-serving tests show up to 40–50% faster time to first token at equivalent context length, or support for context windows up to twice as long at equivalent latency. These are Marvell’s stated results, not independent benchmark findings. Its full announcement has further technical detail.
Our top picks
- Six memory channels
The new controllers add 50% more channels than the previous generation, helping expand memory capacity and bandwidth. - Memory shared across four hosts
Multi-host connectivity is designed to let AI clusters use memory resources more efficiently. - Inline compression
Marvell says LZ4 compression can deliver two to three times more effective capacity, depending on the workload. - Faster LLM responses
The company reports up to 40–50% faster time to first token at equivalent context length. - Roomier context windows
Marvell says the controllers can support context windows up to twice as long at equivalent latency. - Data-centre security features
The release lists post-quantum cryptography, Caliptra 2.0, CXL IDE and AES-256 memory encryption.
Why it matters
AI servers can be limited not only by the processors doing the work, but by how much data they can keep close at hand. CXL memory expansion and sharing aim to give operators more flexibility as models and context windows grow, without treating every workload as a reason to buy a wholly new server. The practical payoff will depend on how these controllers perform in real systems and on the workloads being run.
Our read
The most interesting part is the combination of extra capacity, shared memory and compression, rather than any single headline speed. If the gains hold up beyond Marvell’s own tests, they could make existing infrastructure work harder. The controllers are not yet shipping, though: Marvell expects samples in Q1 2027, so buyers will need to wait for systems and independent testing before judging the claims.
What to watch
- Whether sampling begins on schedule in Q1 2027.
- How the controllers perform on real AI workloads, including longer-context inference.
- Which server makers and memory configurations support the new parts.
Discussion spark: For AI infrastructure, would you prioritise adding memory capacity through CXL, or buying more compute even if memory remains the bottleneck?
Sources and evidence
- Source update (12 October 2026, 00:00 UTC)
not affiliated with or endorsed by Marvell Technology