At the recent Ignite conference, Microsoft announced the Azure Boost DPU, its first in-house DPU designed for efficient, low-power, data-centric workloads. According to the company, this single piece of silicon combines high-speed Ethernet and PCIe interfaces, network and storage engines, data accelerators, and security features into a fully programmable system on a chip.
The Azure Boost DPU originated from Fungible, a DPU fabricator Microsoft acquired in December 2023. Founded by ex-Apple and Juniper Networks engineers, the Fungible team joined Microsoft’s infrastructure engineering division after the acquisition.
The Azure Boost DPU is a hardware-software co-design optimized for Azure, featuring a lightweight data-flow operating system. It delivers higher performance, lower power consumption, and enhanced efficiency compared to traditional systems, with expectations of running cloud storage workloads at three times less power and four times the performance of CPUs. Additionally, DPU-based systems include a custom application layer that utilizes integrated data compression, protection, and cryptography engines, enhancing security and reliability.
(Source: Bluesky post)
GPUs have transformed AI model training but have drawbacks like high power consumption, cooling needs, and current shortages, with Nvidia's GPUs sold out for the following year. Custom accelerators are emerging as alternatives due to better price performance and efficiency. AWS and Google have been using custom chips, like AWS's Trainium, Inferentia, and Google's TPUs. At the same time, Microsoft recently entered the market with its Maia and Cobalt chips, focusing on energy efficiency.
In the Ignite 2024 book of news, the company states:
With Azure Boost DPUs joining the processor trifecta in Azure (CPU—AI accelerator—DPU), enhanced by the hardware security capabilities of Azure Integrated HSM, continued innovations in Cobalt and Maia, paired with state-of-the-art networking, power management, and hardware-software co-design capabilities, Microsoft continues to optimize every layer of its infrastructure in the era of AI.
In addition, Chris Evans, a Storage Architect, posted on Bluesky:
The amount of new silicon developed by Microsoft, AWS, GCP, etc, should be a worry for traditional vendors. It will represent a divergence from traditional standards and diverge the TCO models.
And Shawn Chaucan, a Generative AI Consultant, tweeted:
Microsoft's new chips are a game-changer for AI and data security! The efficiency and performance boosts are impressive. Can't wait to see how they'll impact the cloud computing landscape.
Lastly, by adding DPUs, Microsoft creates a Cloud AI infrastructure, including x86 and Cobalt CPUs, Maia accelerators, and Azure Boost DPUs.
