Marvell photonics module puts HBM memory 50 meters from the GPU at sub-microsecond latency
rohanpaul_ai · x · 2026-10-20
Marvell's Dave Lazovsky demonstrated a photonics module that moves HBM + DDR memory up to 50 meters from a GPU — versus the usual 1.5 mm. Memory no longer must sit next to compute, letting AI clusters share massive capacity without being locked to the accelerator package. Latency stays in the hundreds of nanoseconds, practical for real memory use; memory and compute can now scale on separate axes.
More from Infra
- DOE backs autonomous lab testbeds as grid power limits US AI scale — ingliguori · 2026-10-20
- FlashMap: flow map speculative decoding delivers >6x faster Qwen3-8B inference — kalyan_kpl · 2026-10-20
- NVIDIA touts PyTorch's 50M+ monthly CUDA downloads with a decade of optimizations — NVIDIAAI · 2026-10-20
- NVIDIA invests in 5 UK startups and commits $250M compute for agentic safety research — dr_alphalyrae · 2026-10-20
- Morgan Stanley: $700B may reach Nvidia by 2027, but US data centers face 32GW power gap — rohanpaul_ai · 2026-10-20
- MiniCPM-V 4.7 runs locally at 100 tok/s on just ~3 GB of memory — NielsRogge · 2026-10-20