Reactions from ranked influencers
2 posts
I see that a GPU can receive data in the ~50GB/s range. It would not be simple to make a piece of hardware with e.g. 16 M.2 ports that can directly read them in parallel and feed the GPU? Again, not my field, just curious to understand why the situation is what it is.
I feel it's not that far from possible but doubt default PCIe topologies make it realistic. GPU Direct has allowed direct NVME -> GPU DMA transfers initiated by CPU for awhile now. Need to keep CPU out of the transfers. But usually don't have enough lanes to cover NVME drives + the GPU on the same segment of the PCIe bus. I don't think you'd want to cross over the CPU's root complex or other bus bottlenecks. Also data would have to be organized exactly as you need it on the GPU, otherwise you'd need a (slow) processor or extra logic in GPU or setting between to reorganize... there's been research and work on GPU initiated transfers (google BaM). Not sure how far along that is?