This is part 3 of a series of blogs covering the NVIDIA GTC '22 event. You can read part 1 here and part 2 here.
As workloads evolve, the data center dynamics will shift to accommodate this evolution, with the announcements made at this GTC event, NVIDIA has set out their strategy and highlighted the data center components that they think will need to evolve in step with workloads. The per-instance partitioning introduced with the Hopper GPU architecture shows that multi-tenant sharing of high-end acceleration technology is likely to be an important factor in controlling costs and democratizing such technology. It is likely aimed at the hyperscaler business model but is equally important to large enterprises looking to share high-end acceleration technologies throughout their whole organization and not make technology decisions in isolation.
Further operational cost awareness is demonstrated by both the Grace Superchip and the NVLink interconnect that enables it, NVLink employs ultra-energy efficient components and grace uses low-powerLPDDR5x memory, this results in a 144-core chip that consumes only 500 Watts of power. This design again demonstrates that driving down the operational costs is believed to be a key factor in democratizing AI, which is especially true of the hyperscaler models.
In opening their NVLink interconnect to all their processors it demonstrates that NVIDIA believes these acceleration technologies will play an increasingly important role in meeting the demands of the modern, evolving workload and suggests that more flexibility is expected from solutions. Specialized heterogeneous systems will also play an increasingly important role in servicing future workloads. This trendline is further reinforced by NVIDIA opening their interconnect technology to customers and partners to develop their own solutions that take advantage of the NVIDIA ecosystems whilst also servicing their existing customer relationships and innovating in line with their own strategies. Opening NVLink as announced will dramatically increase the diversity of potential custom systems that can be built for new or specific workloads. This is visionary thinking, by encouraging productive collaborations NVIDIA is signaling that, whilst it wants to control the stack, it remains open to future collaborations and supports innovation in all fields.
The full-stack strategy that NVIDIA has adopted, announced at this GTC event, will help lower the barrier to entry for the enterprise looking to gain a competitive edge by distilling their valuable business data into more valuable actionable data through AI processing. NVIDIA is providing the technology that enables the enterprises of today to become the AI factories of tomorrow. One question remains, with NVIDIA being a full-stack company, will their future GPU Technology Conference events need to be re-named to “Intelligence Factory Technology Conference”?
Part 1: Full Stack, Layered Technology will Democratize AI
Part 2: NVIDIA’s Holistic Approach Marries Technology Components to Create Powerful Union
