In this week’s podcast, Chris and Martin look at NVIDIA’s GPUDirect Storage (GDS), a technology for moving data directly from persistent storage to GPUs, bypassing the CPU. The aim of the technology is to provide greater throughput to keep GPUs active, and looking at some of the thoughts from our podcast with Liqid (#204), it’s clear that this feature is needed. However, with such quick adoption by many storage vendors, is the announcement of GDS more of a marketing exercise? We dig into the details and what the announcement of GDS could mean for the future enterprise.
On a side note, there was indeed a remake of Magnum P.I. in 2018, following on from the original series in 1980. The original was a classic; we’re not sure about the remake.
Elapsed Time: 00:30:42
Timeline
- 00:00:00 – Intros
- 00:01:00 – The Chiacoin price has collapsed!
- 00:02:15 – What is GPUDirect?
- 00:06:05 – GDS has Local and Remote modes
- 00:07:15 – GDS appears to be a data mover offload
- 00:09:05 – High throughput application look to benefit most from HDS
- 00:10:00 – GDS is not a generic replacement for external storage I/O
- 00:11:40 – Many vendors supported the announcement of GDS
- 00:12:53 – What drivers and mods will be need for GDS?
- 00:14:38 – Disaggregated architecture strikes again with ConnectX
- 00:16:04 – GPUs are a next-generation data centre technology
- 00:17:43 – CPU lock-in is a thing – could GPU lock-in exist too?
- 00:19:28 – Will NVIDIA lock Arm and GPU technologies together?
- 00:23:29 – Is the Von Neumann architecture dead?
- 00:25:30 – Will all vendors eventually support GDS?
- 00:30:00 – Wrap Up
Related Podcasts & Blogs
- #204 – Liqid Composable Disaggregated Infrastructure
- #190 – NVIDIA BlueField SmartNICs and DPUs
- Intel Under Pressure as NVIDIA Announces Grace CPU
- Persistent Memory in the Data Centre
Copyright (c) 2016-2022 Unpacked Network. No reproduction or re-use without permission. Podcast episode #h17t.
Podcast (storageunpacked): Play in new window | Download