This week, Chris and Martin chat to Kevin Deierling, SVP of Marketing for Networking Products at NVIDIA. SmartNICs and DPUs (Data Processing Units) are starting to become mainstream as application use-cases such as AI and analytics drive a need for greater data throughput and performance. Kevin explains the design and thinking behind BlueField, NVIDIA’s family of DPU products that combine offloaded network, storage and security functionality.
Elapsed Time: 00:47:36
Timeline
- 00:00:00 – Intros
- 00:01:30 – We’re not a networking podcast!
- 00:02:15 – Why will we need DPUs and SmartNICs?
- 00:04:30 – Von Neumann is diverging
- 00:06:50 – What is the BlueField architecture?
- 00:08:45 – A DPU could act as a storage array controller
- 00:11:30 – Storage DPUs make devices appear local
- 00:13:00 – DPUs enable efficient bare-metal server deployments
- 00:15:00 – Storage, networking & security use around 30% of traditional cores
- 00:16:00 – Does a DPU represent better or worse performance than CPU?
- 00:20:15 – NVIDIA DPUs emulate existing devices, reducing application changes
- 00:25:00 – Where does the outboard management take place?
- 00:26:00 – DOCA is the application framework for DPUs
- 00:28:00 – BlueField 2X combines GPU and DPU on the same card
- 00:31:40 – DPUs enable the real-time nature of data processing
- 00:35:40 – Mainframe reference!
- 00:38:00 – Where will initial adoption take place?
- 00:40:00 – How does the use of SmartNICs affect TCO?
- 00:45:10 – The future is 1000x improvement with BlueField 4
- 00:46:00 – Wrap Up
Related Podcasts & Blogs
- #96 – Discussing SmartNICs and Storage with Rob Davis from Mellanox
- #177 – SmartNICs and Project Monterey
- #180 – SmartNICs – Pliops Storage Processor
- VMware Project Monterey – First Impressions
- Fixing the x86 Problem
Copyright (c) 2016-2022 Unpacked Network. No reproduction or re-use without permission. Podcast episode #ky1r.
Podcast: Play in new window | Download