Nvidia Unveils Scale-In Networking, Blending Front-End and Back-End Traffic

Nvidia's latest networking innovation is called 'scale-in,' which some might consider an unusual addition to the established trio of scale-up, scale-out, and scale-across. Nvidia claims that with its Inference Context Management Storage (ICMS) renamed Context Memory Storage (CMX), it has already broken the three-part rule twice, and now introduces a fourth pillar.

Gilad Shainer, Nvidia's networking chief, explained to SDxCentral that scale-in is not about front-end or back-end traffic but rather a blend of both. He emphasized that traditional north-south traffic models no longer apply in the AI factory context, where multiple technologies come into play, including Astra (Advanced Secure Trusted Resource Architecture) for out-of-band management and policy enforcement.

Shainer highlighted the capabilities of BlueField-4, which extends security and orchestration control directly into scale-out and scale-across fabrics as well as north-south traffic. He noted that this represents an 18-fold increase in secured traffic managed compared to previous generations.