Marvell Announces Innovations in AI Memory

<p><b>SANTA CLARA<&sol;b> – <a href&equals;"http&colon;&sol;&sol;www&period;marvell&period;com&sol;">Marvell Technology<&sol;a>&comma; a provider of data infrastructure semiconductor solutions&comma;  announced new innovations across its AI memory infrastructure portfolio&comma; advancing its position spanning server-level AI storage&comma; rack-scale CXL memory expansion and pooling&comma; and pod-level optical shared memory&period; The portfolio helps hyperscalers and cloud providers scale memory more independently from compute&comma; improving infrastructure utilization&comma; scalability and token efficiency for agentic AI inference&period;<&sol;p>&NewLine;<p>As agentic AI inference scales&comma; memory capacity&comma; bandwidth and connectivity are becoming as critical as compute&period; Larger models&comma; longer context windows and growing KV caches are driving unprecedented memory demands and exposing the limits of traditional server-attached architectures&period; Increasingly&comma; AI performance and token efficiency depend on how efficiently processors can access and move data across memory resources&period;<&sol;p>&NewLine;<p>Memory disaggregation addresses this challenge by enabling memory to be expanded&comma; pooled and shared more independently of compute&period; By making memory more accessible and reducing data movement and latency&comma; the Marvell portfolio helps reduce GPU stalls and improve model FLOPs utilization&comma; allowing AI infrastructure to generate more tokens within the same data center footprints and power envelopes&period;<&sol;p>&NewLine;<p>&OpenCurlyDoubleQuote;AI infrastructure is moving beyond isolated servers to systems where compute&comma; memory and connectivity operate seamlessly together&comma;” said Will Chu&comma; executive vice president and general manager&comma; Custom Cloud Solutions&comma; at Marvell&period; &OpenCurlyDoubleQuote;As AI scales&comma; memory must scale more independently of compute so resources can be deployed where they deliver the greatest value&period; With the industry’s broadest AI memory infrastructure portfolio&comma; Marvell is helping customers improve utilization&comma; boost token efficiency and scale AI without compromising performance&comma; power or cost&period;”<&sol;p>&NewLine;<p>&OpenCurlyDoubleQuote;As AI workloads grow larger and more complex&comma; memory capacity&comma; bandwidth&comma; latency and data movement are becoming primary constraints on AI performance&comma;” said Alan Weckel&comma; co-founder and technology analyst at 650 Group&period; &OpenCurlyDoubleQuote;Marvell’s memory and storage portfolio gives hyperscalers and cloud providers a strong foundation for building scalable&comma; efficient AI systems capable of supporting increasingly advanced workloads&period;”<&sol;p>&NewLine;<p><b>Server-level AI Storage<&sol;b><&sol;p>&NewLine;<p>The Marvell Bravera SC6 PCIe® 6&period;0 SSD controller advances storage performance for AI inference&comma; enabling more KV cache to move from high-bandwidth memory to SSD to improve infrastructure efficiency&period; Doubling the performance of the widely deployed Bravera SC5 PCIe 5&period;0 SSD controller&comma; Bravera SC6 helps cloud providers support demanding AI&comma; cloud and enterprise workloads while improving performance&comma; reducing write amplification and extending NAND endurance&period; Optimized for hyperscale deployments&comma; Bravera SC6 supports NAND from multiple suppliers&comma; providing cloud providers with greater flexibility in storage sourcing and deployment&period;<&sol;p>&NewLine;<p><b>Rack-level Memory Expansion and Pooling<&sol;b><&sol;p>&NewLine;<p>The Marvell Structera X memory expansion solutions help hyperscalers scale memory more efficiently for increasingly memory-intensive AI workloads&period; Developed in close collaboration with leading hyperscalers&comma; Structera X enables customers to expand&comma; optimize and better utilize memory resources&comma; including extending the value of existing memory investments while supporting demanding AI inference applications&period; As larger models&comma; longer context windows and growing KV-cache requirements make memory capacity and bandwidth critical infrastructure constraints&comma; Structera X provides a path toward more flexible and disaggregated memory architectures&period; By enabling larger memory pools and more efficient sharing of resources across servers&comma; the platform helps improve infrastructure utilization&comma; operational efficiency and total cost of ownership while laying the foundation for future advances in CXL-based memory pooling and sharing&period;<&sol;p>&NewLine;<p><b>Pod-level Optical Shared Memory<&sol;b><&sol;p>&NewLine;<p>The Marvell Photonic Fabric memory modules&comma; Photonic Fabric NIC and Photonic Fabric chiplets are foundational elements of a multi-rack optical shared-memory architecture for AI inference&period; The solution creates a new shared-memory tier across multiple XPUs and racks up to 50 meters&comma; enabling up to 32TB of warm KV cache offload with high bandwidth and extremely low latency&period; By expanding memory access across AI clusters and enabling KV cache to be loaded from a shared memory tier rather than storage&comma; Photonic Fabric helps increase inference throughput&comma; support larger models and longer context lengths&comma; and increase token efficiency by delivering up to 2-3x higher token throughput within existing data center footprints and power envelopes&period; <b><&sol;b><&sol;p>&NewLine;<p>The Marvell Bravera SC6 PCIe 6&period;0 SSD controller is expected to begin sampling in Q4 2026&period;<&sol;p>&NewLine;

Editor

Knit Health Lands $35 Million Led by NEA

SAN FRANCISCO -- Knit Health, a healthcare AI company building a foundational model to predict…

7 hours

Bedrock Robotics Raises $600 Million Series C

SAN FRANCISCO -- Bedrock Robotics has raised a $600 million Series C financing at a…

7 hours

Zipline Scores $1.5 Billion Investment

SOUTH SAN FRANCISCO — Zipline has raised $1.5 billion at a $21.5 billion post-money valuation,…

3 days

Jensen Huang Out as NVIDIA CEO

NVIDIA founder Jensen Huang has resigned Monday as President, Chief Executive Officer, and a member…

3 days

Syncere Scores $11 Million Pre-Seed Round

PALO ALTO -- SYNCERE, a home robotics startup, has scored $11 million in pre-seed funding…

2 weeks

Amazon Offers Members Food Delivery From Uber Eats

SEATTLE & SAN FRANCISCO -- Amazon is teaming up with Uber to offer food delivery…

2 weeks