In 2013, Amazon Web Services introduced its C3 instance type, vaguely hinting at ‘enhanced networking’ enabled by Intel Virtual Function interfaces. This enhancement involved network cards capable of running workloads such as firewalls and load balancers alongside traditional packet processing. AWS found this beneficial as it freed server resources, while improving network speed and security. Other hyperscalers also developed similar SmartNICs, or DPUs, with Mellanox transforming the idea into ‘BlueField’ in 2017, enhancing data speeds across flash storage networks.

VMware noticed the potential and began adapting its hypervisor to run on these devices, while Nvidia acquired Mellanox in 2019. Intel entered the scene calling their solution ‘infrastructure processing units’ in 2021, followed by AMD’s acquisition of Pensando in 2022 to join the DPU ecosystem. Despite backing from various tech giants, the widespread deployment of SmartNICs remained limited.

The majority of the DPU market still exists within service providers that leverage server capacity. Innovations such as DPU-enabled Ethernet switches are being developed, promising to broaden the appeal and usage of DPUs. Industry leaders like Cisco have integrated DPUs into their product strategies with systems such as the ‘Hypershield’ security solution and smart switches.

Artificial Intelligence may provide a much-needed boost to DPU deployment. Gartner and Nvidia’s reference architectures for AI position DPUs as essential components, while Red Hat supports DPUs in its OpenShift platform, expecting them to optimize various workloads and play a significant role in AI-driven tasks. The evolving AI landscape could finally usher in the DPU era.