Marvell展示AI服务器光互连共享内存模块
Marvell展示了Photonic Fabric Memory Module,这是一款通过光互连把共享内存与处理器分离并连接至AI服务器的设计。16个模块组成的设备计划提供32TB DDR5和1,152GB HBM3e,最多供16个主机共享;单模块设计包括72GB HBM3e、最多8个DDR5 DIMM插槽、Photonic Fabric ASIC及7.2Tbps光I/O。该设计仍处于预硅阶段,尚无工作硬件实测吞吐量,目标于2027年向客户送样。
Marvell称共享内存层可支持跨节点KV缓存,并让内存容量独立于加速器扩展,但TechRadar指出成本节省和性能尚未量化。
- Marvell reveals Photonic Fabric Memory Module, the building block of a 16-module appliance sharing 32TB of DDR5 and 1,152GB of HBM3e across 16 hosts over fiber
- The design dates back to Celestial AI's 2025 presentations, currently remains pre-silicon, and is targeting customer sampling by 2027
- Nvidia is already hedging with a $2 billion Marvell stake in addition to its own BlueField-4 STX context memory, Enfabrica's pooling tech, and a co-packaged optics roadmap
Marvell has put physical hardware behind one of the more ambitious ideas in AI infrastructure.
In a video posted to X, and earlier in a theCUBE interview recorded around its October 6 Investor Day, the company showed off its Photonic Fabric Memory Module, which it bills as the "first optically interconnected high-capacity, high-bandwidth memory."
The pitch is simple: use light instead of copper so that dozens of terabytes of memory can sit in a separate box, racks away, and be shared by many processors at once, something that could effectively birth the widespread use of communal or shared memory across multiple AI servers.
Two terabytes at a time
Some clarity, however, is warranted: the 32TB figure belongs not to the module but to the Photonic Fabric Memory Appliance, which gangs 16 modules together and shares the pool across up to 16 hosts.
#Marvell#Photonic Fabric Memory Module#Photonic Fabric Memory Appliance#Photonic Fabric ASIC#Celestial AI#Nvidia#BlueField-4 STX#Enfabrica#EMFASYS#NVLink Fusion