# GPUDirect?

**URL:** <https://discourse.julialang.org/t/gpudirect/46424>\
**Category:** Package Announcements\
**Tags:** gpu, cuda, big-data\
**Created:** [September 11, 2020, 3:33am UTC](https://discourse.julialang.org/t/gpudirect/46424 "2020-09-11T03:33:25Z")\
**Posts on this page:** 5\
**Page:** 1

<div class="post-metadata">

**Author:** ![Raf](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/raf/32/3383_2.png) [@Raf](https://discourse.julialang.org/u/Raf)\
**Post date:** [September 11, 2020, 3:33am UTC](https://discourse.julialang.org/t/gpudirect/46424/1 "2020-09-11T03:33:25Z")

</div>

Has anyone experimented with direct transfer using nvidias GPUDirect?

> **[GPUDirect Storage: A Direct Path Between Storage and GPU Memory | NVIDIA...](https://developer.nvidia.com/blog/gpudirect-storage/)**
>
> As AI and HPC datasets continue to increase in size, the time spent loading data for a given application begins to place a strain on the total application’s performance. When considering end-to-end…

It would be great to have a direct HDF5 to GPU pipeline.

---

<div class="post-metadata">

**Author:** ![xiaodai](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/xiaodai/32/15937_2.png) [@xiaodai](https://discourse.julialang.org/u/xiaodai)\
**Post date:** [September 11, 2020, 7:31am UTC](https://discourse.julialang.org/t/gpudirect/46424/2 "2020-09-11T07:31:03Z")

</div>

some talk about it on gpu channel in slack. heards it’s only for windows and design for games for now

---

<div class="post-metadata">

**Author:** ![Raf](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/raf/32/3383_2.png) [@Raf](https://discourse.julialang.org/u/Raf)\
**Post date:** [September 11, 2020, 7:38am UTC](https://discourse.julialang.org/t/gpudirect/46424/3 "2020-09-11T07:38:29Z")

</div>

I found that thread, there are just a bunch of sources saying its a lot more than that:

> **[s9577-rapids-the-platform-inside-and-out.pdf](https://developer.download.nvidia.com/video/gputechconf/gtc/2019/presentation/s9577-rapids-the-platform-inside-and-out.pdf)**
>
> 32.33 MB

> **[Nvidia GPU Accelerators Get A Direct Pipe To Big Data - The Next Platform](https://www.nextplatform.com/2019/08/08/nvidia-gpu-accelerators-get-a-direct-pipe-to-big-data/)**
>
> Nvidia has unveiled GPUDirect Storage, a new capability that enables its GPUs to talk directly with NVM-Express storage. The technology uses GPUDirect’s RDMA facility to transfer data from flash storage into the GPU’s local memory without needing to...

---

<div class="post-metadata">

**Author:** ![rkube](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/rkube/32/211198_2.png) [@rkube](https://discourse.julialang.org/u/rkube)\
**Post date:** [November 12, 2022, 12:05pm UTC](https://discourse.julialang.org/t/gpudirect/46424/4 "2022-11-12T12:05:52Z")

</div>

Hi, reviving this thread since I’m also interested in this. Some of my ML workflows with big datasets are I/O limited. I found this presentation [https://www.hdfgroup.org/wp-content/uploads/2020/10/GPU\_Direct\_IO\_with\_HDF5-\_John\_Ravi.pdf](https://www.hdfgroup.org/wp-content/uploads/2020/10/GPU_Direct_IO_with_HDF5-_John_Ravi.pdf) but couldn’t find an implementation.

---

<div class="post-metadata">

**Author:** ![ToucheSir](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/touchesir/32/14411_2.png) [@ToucheSir](https://discourse.julialang.org/u/ToucheSir)\
**Post date:** [November 12, 2022, 5:21pm UTC](https://discourse.julialang.org/t/gpudirect/46424/5 "2022-11-12T17:21:22Z")

</div>

The relevant CUDA.jl issue is [Support for GPUDirectStorage · Issue #1630 · JuliaGPU/CUDA.jl · GitHub](https://github.com/JuliaGPU/CUDA.jl/issues/1630#issuecomment-1292099494). Seems like it’s PRs welcome.