SDxCentral – NVIDIA and Fungible Dispel DPU Misconceptions
The latest article by SDxCentral, “NVIDIA and Fungible Dispel DPU Misconceptions”, does a great job of describing some of the overall benefits of the data processing unit (DPU), such as efficiency and lower overall cost. It also covers some of the differences between the architectures of the NVIDIA® BlueField® DPU and the Fungible DPU™, but it stops short of explaining some of the key differences, and why that matters for enabling truly data center scale efficiency.
The DPU accelerates input/output workloads. This is true for both NVIDIA and Fungible DPUs, but the two companies are aimed at solving very different problems in the data center. NVIDIA has made an SDK (software development kit) available so that users can develop software on their DPU. There are developers lined up to do so, but the resulting software will only target that user’s specific solutions. It will accelerate one process. There is value in that, but only for that one user. When the problem changes, the code needs to be rewritten, each and every time.
Fungible’s DPU is also programmable but was designed from scratch to handle a broad array of data-centric problems, including storage, networking, security, and more. The Fungible DPU’s unique architecture was created to allow programming pipelines of operations for a variety of tasks, without doing any look-aside offloads to general-purpose processors. The benefit here can be tremendous in terms of performance and latency improvements, as well as being able to address a much broader set of problems. For example, a storage workload may need to handle NVMe over TCP offload, encryption, compression, and erasure coding. All of these operations can be handled as data flows through the DPU, moving through each accelerator as required to process the prescribed pipeline, all at line rate. The next workload could involve offloading TLS acceleration and layer 2 networking functions, which will follow a different pipeline to execute these operations. In fact, both of these workloads can happen at the same time, multiple times, all at line rate, without having to take a round trip off to a general-purpose processor.
In contrast to NVIDIA, Fungible provides a complete solution that is ready to go. We’ve done the programming and the hardware integration work. We ship a complete NVMe/TCP all-flash, scale-out storage cluster, which scales in performance AND efficiency as the cluster grows in scale. It’s a complete, bare-metal virtualization and orchestration system that allows dramatic simplification of your server fleet while consolidating all storage to drive efficiency and lower costs. In addition, our systems install in standard racks, connect to standard ethernet networks, and enable the benefits of composability. And there is MUCH, MUCH more to come in 2022!
As users work to build their next-generation data centers to be more agile, efficient, and cost effective to handle the exponential growth of data and applications, Fungible is delivering solutions to get you all the way there, not just point solutions that require constant tweaking and stitching together. We are working to bring the cloud experience back to the data center. The Fungible DPU sits at the center of this work, as a key enabler, while we build efficient and scalable solutions to radically transform and modernize data center infrastructure. We welcome you to join us on this journey.