Worth noting that Infiniband RDMA and NVMeoF exists & is pretty similar to this proprietary GPUDirect Storage idea. Now that Nvidia owns Mellanox though, who th

Worth noting that Infiniband RDMA and NVMeoF exists & is pretty similar to this proprietary GPUDirect Storage idea. Now that Nvidia owns Mellanox though, who the frak knows what the rest of the damned planet is going to do. Fart-balls! That said, huge respect to Nvidia for chasing the puck/seeing the trend & bringing it to developers. This CuDF framework could not be better targeted: > cuDF is a Python GPU DataFrame library (built on the Apache Arrow columnar memory format) for loading, joining, aggregating, filtering, and otherwise manipulating data. cuDF also provides a pandas-like API that will be familiar to data engineers & data scientists, so they can use it to easily accelerate their workflows without going into the details of CUDA programming. Taking best of breed data-storage, Apache Arrow, and building a python library that hardware optimizes the storage<->gpu flow around it, is some excellent marksmanship. It'll be interesting to see what Nvidia does next in this area. Nvidia is stuffing BlueField smart-fabric NICs into all their new Grace systems[1]. First less obvious would be some push-down optimization, do some filtering of the data coming off the NVMe at the NIC itself. Amazon meanwhile has Nitro, which is a brilliant way to let cpus see disaggregated devices as though they were native/close, which bypasses the need for RDMA, kind of is an RDMA-like but in hardware system. There's not a lot on my radar I can identify that can counter this, since Infiniband got left to rot. There's RoCE (RDMA over Converged Ethernet) but basically the mainstream market has forefit being good. Maybe some day CXL can be useful for GPUs? [1] https://nvidianews.nvidia.com/news/taiwans-tech-titans-adopt-worlds-first-nvidia-grace-cpu-powered-system-designs https://nvidianews.nvidia.com/news/taiwans-tech-titans-adopt...