Computing Infrastructure Optimized for the Demands of AI
2020 is barely one-third over and it’s already one of the great outlier years we may ever experience. But even that extreme level of uncertainty can’t stop us from looking ahead to examine the forces, trends and big ideas at work in the data center sector.
Here are five themes that will define data center AI infrastructure in 2020:
1. Enterprises will focus on building GPU clusters with ultra-fast storage. At the head of the list of solutions designed to meet the most rigorous standards for performance and speed are NVMe all-flash arrays and servers, which offer the following benefits:
- Optimization for maximum capacity or highest throughput performance and application responsiveness
- Broad support of advanced form-factors including EDSFF, U.2, and NF1
- Easy capacity expansion
- Tailoring for ideal data center, data warehousing, software-defined storage and hyperscale/hyperconverged performance
Enterprises that that build their GPU clusters using NVMe all-flash arrays and servers have the inside track on delivering the advanced computing
environments required to enable the highest level of data center AI performance.
2. Hardware manufacturers will rise to meet the infrastructural challenges of AI. A hardware arms race is emerging in response to the explosion of data center AI demand, creating an environment that is prioritizing chip innovation. Intel and NVIDIA recently debuted cutting-edge hardware for the high-performance computing community. Their newest GPUs are targeted above the typical enterprise computing space and being applied to specialized workloads, like data-crunching for artificial intelligence.
New entrants include Cerebras Systems, which is turning heads with a system that squeezes 400,000 compute cores into a 15U rackmount chassis, and Groq, which boats a new chipset capable of one quadrillion operations per second.
3. Cooling challenges will create a potential AI deployment bottleneck. Those eye-popping specs don’t come easily and developing new cooling technologies to cope with the skyrocketing chip performance and rack density requirements of advanced AI deployments will be vital. If all this new AI gear is going to see widespread use, direct-to-chip liquid cooling and immersion solutions will need to become more advanced, efficient and affordable.
Liquid cooling is already demonstrating gains, but there is significant work yet to be done. Advances in these technologies and processes will go a long way toward defining data center AI infrastructure in the near and intermediate future.
4. AI development will become industrialized. Throughout the sector, there is a distinct lack of process standardization and mechanization in AI development. The journey from conceptualization through prototyping and training, to deployment and steady state is notoriously artisanal, evolutionary and fickle from one enterprise to the next. Shifting to a more industrialized model has been underway for several years.
Organizations are prioritizing the intense analysis of workflows to deliver processes that are more dependable, efficient and replicatable. Systems integration is becoming increasingly automated and itself managed by machine-learning applications, and IT teams are refocusing to rapid iteration and innovation. More advanced workflow industrialization will unlock the potential for AI to transition into the upper tier of enterprise IT-managed workloads.
5. Power infrastructure will be increasingly green. Advances in capacity and throughput require more power than ever and data center infrastructure funds are prioritizing green solutions. Future data centers will be powered by solar, wind and hydro, use assets that include waste-to-energy or carbon capture technologies, and find other ways to ensure that the demand to bring new resources to deploy IT capacity at scale has a reduced impact on the environment.
Macquarie Group, one of the world’s largest infrastructure investors, is determined to “amplify climate mitigation and adaptation.” Cologix, a leading provider of colocation and interconnection services that is part of the sprawling Stonepeak Infrastructure Partners group of companies, is likewise providing leadership in the move to renewable energy for data center applications.
Transformation of processes, technologies and practices is occurring across the data center AI infrastructure landscape. Savvy stakeholders are planning now for the emergent factors that will define tomorrow’s market.