Marvell 的 Dave Lazovsky 展示了一个光子模块，可将 HBM + DDR 高带宽内存搬到距 GPU 最远 50 米处，而传统方案通常只有 1.5 mm。该方案让内存首次不必紧贴计算芯片，AI 集群可共享大容量内存而不被加速器封装锁定，延迟维持在数百纳秒级，足以支撑实际内存使用。内存与计算由此可以各自独立扩展。
Marvell’s Dave Lazovsky is holding a tiny module that uses photonics to move high-bandwidth memory (HBM + DDR) as far as 50 meters away from the GPU.
this can sit up to 50 meters from a GPU instead of the usual 1.5 mm.
For the first time, memory no longer has to live right next to compute, letting AI clusters share massive capacity without being locked to the accelerator package.
Latency stays in the hundreds of nanoseconds, close enough for practical memory use.
Basically, Memory and compute now scale on separate axes instead of being locked together. https://x.com/MarvellTech/status/2112227641759871266/video/1
来源：Rohan Paul · x.com