NVIDIA has several announcements at GTC Fall 2021. The first one that we are going to cover is the NVIDIA Quantum-2 and ConnectX-7 400Gbps era InfiniBand products. This is increasing the overall bandwidth of InfiniBand networks to help scale out cluster performance.
NVIDIA Quantum-2 400G Switches and ConnectX-7
NVIDIA Quantum-2 is based around 400G NDR InfiniBand. Along with the raw switching speed, NVIDIA is also adding more processing to the Quantum-2 switches for additional performance when scaling workloads across cluster nodes.
The Quantum-2 is a large TSMC 7nm chip that can handle up to 64-ports of 400G or 128-ports of 200G connections.
Of course, NVIDIA needs more than just the switches to make 400G NDR InfiniBand possible For that, there is the NVIDIA ConnectX-7 family that can serve as endpoints for Quantum-2 switches.
The other option is the next-gen BlueField DPU, the NVIDIA BlueField-3 DPU. We have seen a lot on BlueField-3, but it is somewhat surprising this is not sampling until May 2022. Both Intel and Marvell are aggressively pushing in the DPU space and so this is an area where we were hoping that NVIDIA would be shipping production parts by May 2022, not just sampling.
NVIDIA does have its new DOCA 1.2 framework. We were also a bit surprised that there are only 1400 DOCA developers. NVIDIA is also focusing a lot here on security rather than some of the more infrastructure management that we are seeing from cards like the Intel Mount Evans and the AWS Nitro series.
NVIDIA is leaning into the security space making the BlueField a component for zero-trust security solutions. NVIDIA has several partners here, but there are some big security and networking companies here and only 1400 DOCA developers.
What is more here, NVIDIA has more BlueField ecosystem companies, but the focus right now seems to be on security.
Overall, we are still waiting for OpenStack and similar offerings to support BlueField natively, hopefully this happens before others like Intel take that space.
Something that was emphasized in the keynote was the nanosecond precise timing on Quantum-2 in order to keep clustered nodes synchronized.
Overall, NVIDIA Quantum-2 and ConnectX-7 are a fairly predictable speed bump. We have already gotten hands-on with a 400GbE switch, so InfiniBand needs to continue evolving, and Quantum-2 will be a big bump.
On the BlueField side, the DPU is a technology we are very excited about, but from what we are seeing elsewhere in the market, NVIDIA’s focus on enterprise software partners for zero-trust security with the solution is seemingly more narrow than we were hoping from the products. Still, there is a lot of time until BlueField-3 is out.