NVIDIA Is Working On Vulkan Support With RDMA Memory

Well this will be interesting to see what NVIDIA use-case pans out... NVIDIA engineers are working on a Vulkan extension for making use of RDMA memory.

Remote Direct Memory Access (RDMA) for zero-copy networking with high throughput and low latency is very common for cluster computing and other enterprise scenarios to allow direct memory access from one computer to another without the intervention of the CPU. NVIDIA now though is preparing to support RDMA memory usage in the Vulkan context.

This morning one of their engineers opened a GitHub pull request on the Vulkan-Docs repository for reserving extension 364. When it was first opened it said this extension is for VK_NV_rdma_memory.

But since first noticing that rather interesting RDMA memory extension and getting to writing about it, the pull request was updated to just read: Reserve extension 364. The reservation change now just refers to it as "VK_NV_extension_364" and no reference to VK_NV_rdma_memory. A comment also recommends the pull instead be opened on The Khronos Group's private Gitlab instance.

The extension spec wasn't part of the original pull request with this just about reserving the extension number. So we'll need to wait a bit longer before they go ahead and get this VK_NV_rdma_memory extension included in Vulkan to learn more about NVIDIA's intentions with Vulkan and RDMA memory usage.

At a lower-level than Vulkan, Intel's open-source engineers have been looking at DMA-BUF support for RDMA.