Much has been said about the way artificial intelligence (AI) will alter the networking landscape, usually in reference to its ability to enhance traffic management, analytics and the provisioning of virtual resources.
But AI will also have a pronounced impact on network infrastructure and in particular the fundamental way in which networking silicon processes packets and data streams. The rub is that these changes will happen not because of the new capabilities that AI brings to the table but because AI workloads have vastly different networking requirements than traditional data.
According to Tractica, AI will likely produce a boom in hardware sales that will drive the market from today’s $3.5 billion to $115 billion by 2025. Initially, much of this activity will come in the form of acceleration hardware on the compute side of the data environment, but both storage and networking are likely to see major innovations going forward in order to optimize AI workloads end-to-end. It is also likely that hardware constructs will become more specialized for individual AI workloads, particularly as applications become more adept at provisioning their own resources.
For the network, the key differences between “normal” and AI workloads lie in their dynamism and scale, says HPE’s Thomas Goepel. Traditional data is structured and static, which is fine when offloading to discrete storage and processing pools for intense — and often time-consuming — analytics. AI is more free-flowing and subject to parallel processing technologies like Hadoop. It also requires real-time transit speeds and on-the-fly integration of multiple streams from widely distributed end points. As such, AI will require less point-to-point connectivity and more in the way of flexible, composable fabrics capable of auto-discovery and support for highly abstract architectures. (Disclosure: I provide content services for HPE.)
Clearly, in this world, latency is the enemy. While many organizations are rightfully increasing overall bandwidth from 10 to 25 to 100 Gbps and beyond, a wider pipe alone will not necessarily improve AI performance. Today’s analytic workloads often consist of millions of tiny files that require advanced network adapters capable of meeting massive numbers of I/O requests at latencies as low as 200 microseconds. A company called WekaIO is targeting this need with a new Ethernet fabric technology that leverages Intel’s Data Plane Development Kit (DPDK) to processes twice the amount of data as a local NVMe solution.
AI workloads are also likely to produce a rise in field programmable gate arrays (FPGAs) on the network and elsewhere. Xilinx recently rolled out a new adaptive compute acceleration program (ACAP) called Versal that combines networking, memory and software development tools to address key workloads like AI, IoT and 5G. The idea is to support a wide range of architectures under the same basic framework, utilizing the FPGA’s programming capabilities to optimize the environment after hardware has been deployed. The platform also includes an embedded digital signal processor capable of supporting high-precision floating point workloads while still maintaining an efficient power envelope.
All of this adds up to a network that must be dramatically faster, more versatile and more adaptable than what we have today. More so than previous forms of computing, AI requires broad optimization across disparate resources. If the network cannot provide the connectivity sufficient to this challenge, the entire transition to a more intelligent, autonomous data ecosystem will grind to a halt.
At the moment, much of the attention is on how AI will help the network. Going forward, we’ll have to start thinking harder about how the network can help AI.
Arthur Cole is a freelance journalist with more than 25 years’ experience covering Enterprise IT, telecommunications and other hi-tech industries.