Marvell Pushes Data Centers to Buy AI Memory and Compute Separately
shashib · x · 2026-08-10
Marvell argued at the Flash Memory Summit that hyperscalers should decouple AI memory purchases from compute. As agentic workloads cause key-value caches to swell with context length, server-attached memory has become the primary bottleneck leaving GPUs idle.
To solve this, Marvell introduced a three-tier product lineup:
- Bravera SC6: A PCIe 6.0 SSD controller that offloads KV cache to flash, freeing up HBM.
- Structera X: Expands memory at the rack level using the CXL interconnect standard.
- Photonic Fabric: An optical layer enabling processors to share up to 32TB of memory across racks up to 50 meters apart.
This architectural separation will fundamentally change capacity planning and rack designs for hyperscalers.
More from Infra
- Stop Running Blind: Open-Sourcing specspecs for Speculative Decoding Observability — HamelHusain · 2026-08-10
- Sony and TSMC to Invest $6.3B in Advanced Image Sensor Plant in Kumamoto — pstAsiatech · 2026-08-10
- Running Video Generation on RTX 4060 Ti: Qwen3-vl + MiniMax-H3 Takes 26 Minutes — LuisaPinguinnn · 2026-08-10
- DeepSeek V4 Flash Clears All 22 Coding Tasks on Dual DGX Spark Cluster — AccBalanced · 2026-08-10
- Local LLM Deployment Costs $10K, Taking 24 Years to Break Even vs API — TheZachMueller · 2026-08-10
- WinterMix: New 3-bit MLX Quantization Beats GGUF in Long Context — WinterCharm · 2026-08-10