Nvidia does cybersecurity?!?.....and networking?

About this video

Chuck interviews Kevin Deierling (named in the transcript as SVP of networking at Nvidia, who joined via the Mellanox acquisition) about Nvidia’s move beyond GPUs into networking and cybersecurity. Deierling says Nvidia is the largest provider of networking at 25 gig and above, and introduces the BlueField DPU (data processing unit) as a third pillar alongside CPU and GPU: the data center is the new unit of computing. The problem, he argues, is that software-defined networking, storage and security run badly on CPUs — cloud providers can burn 30 to 40% of CPU cores just running infrastructure they could otherwise sell. The DPU offloads, accelerates and isolates those tasks with hardware accelerators and RDMA, moving files at 100 or 200 gigabits per second with zero CPU. With Palo Alto Networks using the DOCA software platform, he claims firewall deep packet inspection that brought a CPU to its knees at about 10 gig on a 100 gigabit link reached close to 100 gigabits, nine times better, with zero CPU load. Other topics: Morpheus, which uses DPU telemetry plus AI to spot anomalous behavior in logs and traffic rather than fixed rules; Chronos/packet pacing for precise streaming delivery; GeForce Now cloud gaming; Project Monterey putting VMware on BlueField; SMPTE 2110 video production over Ethernet; and Omniverse digital twins, including an AI-generated digital twin of Nvidia’s CEO and his kitchen in a GTC keynote and a BMW factory simulation.

Key takeaways

• Nvidia pitches the DPU alongside the CPU and GPU: the DPU runs software-defined networking, storage and security with hardware acceleration instead of consuming host CPU.
• Deierling claims infrastructure can consume 30 to 40% of a cloud provider’s CPU cores, and that DPUs can move data at 100-200 Gbps via RDMA with zero CPU.
• The Palo Alto Networks example claims nine times better firewall inspection performance, approaching 100 gigabits, by offloading to the DPU via DOCA.
• Morpheus pairs DPU telemetry with AI to flag anomalous behavior (e.g. machines behaving like humans) rather than relying only on fixed log rules.
• DOCA is positioned as the DPU equivalent of CUDA for GPUs, intended to keep partner applications working across BlueField 2, 3 and 4.

Episode details: published November 2, 2021 · 49:22 long. Summary written by Self Growth Videos from the episode itself. Business results and any income figures mentioned are NetworkChuck’s own claims, shared for education — they are not typical or guaranteed.

← Previous: do you need to be good at MATH to learn Python? // Python RIGHT NOW!! // EP 3 · Next: you need to learn Google Cloud RIGHT NOW!! →

All NetworkChuck videos · NetworkChuck profile · NetworkChuck on YouTube