Building a Category-Defining Technology Company: From Foundational Technology to Breakthrough Products to Commercial Success
Pradeep Sindhu, Co-Founder, Executive Chairman and Chief Development Officer
From inception, Fungible’s exclusive focus has been on building the foundational silicon and software technology to revolutionize the performance, economics, reliability, and security of data centers at all scales and geographies. In this journey, our North Star has always been the acknowledgment of two macro trends: First, modern applications are increasingly data-centric, in that the work required to move and store data is as important as the arithmetic and logical operations performed by them. Indeed, in a scale-out architecture, storing and moving data are integral to “computation” at the data center level—something that is equally true at lower levels in the hierarchy of computing elements. Second, the historical improvements that came from advances in silicon technology, the so-called “Moore’s Law advances”, have been slowing significantly, making the approach of focusing primarily on arithmetic and logic operations untenable in the long term. The combination of these two trends literally mandated the invention of a new fully programmable building block specialized to perform data-centric computations efficiently. The self-imposed requirement of full programmability came from the consideration that any fixed hardware approach, while performing a narrow set of computations well, would rapidly become obsolete as new and more demanding data-centric workloads needed to be addressed.
Thus was born the DPU, a term we invented back in 2017 while the industry was still focused on incrementally evolving NICs to be “Smart” and in providing lookaside offloads for a few functions that were too expensive for CPUs to perform. Fungible’s DPU architecture was developed from a clean sheet based on fundamental considerations of what type of micro-architecture it would take to process data-centric workloads in the most general way possible while providing the highest possible performance. Significantly, we did not start by evolving an existing approach. This would have imposed unnecessary constraints and resulted in only incremental gains. Subsequent experience with “Smart NICs” by customers hungry for solutions has shown this in a clear and compelling way: whether one executes a data-centric workload on an X86 core or an ARM core, the result is roughly the same—inefficient execution, as revealed by the number of instructions per clock (IPC) delivered by these cores staying well below 0.25. In sharp contrast, the Fungible DPU™ can deliver a sustained IPC of more than 1.5 per core on data-centric workloads. This is the advantage of starting with a clean sheet and focusing on the fundamentals of a problem rather than following an incremental approach. It is gratifying to see industry leaders finally acknowledge that there is indeed a problem to be solved here, and also take the steps to move in the direction we embarked on five years ago, at least with respect to the language they are using. This is a great endorsement of Fungible’s vision and approach.
In parallel with our silicon effort, we have co-developed FunOS™ a purpose-built Operating System for the Fungible DPU that supports a powerful programming paradigm for writing data-centric applications in standard ANSI-C. On top of FunOS, we have built complete software stacks for storage and networking that are used to power the industry’s most powerful scale-out storage system, the Fungible Storage Cluster. An elastic block store is widely acknowledged to be a hard problem to solve. Yet, starting from scratch, our storage team has delivered the world’s fastest, fully-featured elastic block store on standard NVME over TCP in record time. This speaks not only to the team’s capability but to how easy the Fungible DPU is to program.
Complementing the Fungible Storage Cluster, our FC cards enable CPUs and GPUs to be efficiently disaggregated at scale, enabling the pooling of virtually all data center resources. When used as a storage initiator, a single FC card will deliver a sustained 2M 4K IOPS to applications running on its server while consuming no X86 cores for transport.
Finally, our Composer software enables us to deliver our complete vision of Fungible Data Centers to customers from edge locations: we can create highly customized virtual data centers in mere minutes from underlay server hardware powered by the Fungible DPU while providing the highest levels of performance for specific workloads. To protect our breakthrough technology, we have filed over 60 patents, many of which are fundamental in nature. Any one of these developments would be a heavy lift, not just for a startup but for any company. Taken together, these developments are a staggering achievement and a testament to the dedication and skill of our team.
The Fungible DPU has made two fundamental improvements to the state of the art in data center computing: First, it can perform data-centric computations far more efficiently than CPUs or GPUs while being fully programmable. For example, our storage system based on dual-socket DPUs with 24 NVMe SSDs delivers >15M 4K IOPS over the network using standard NVME over TCP, compared with <1M 4K IOPS for an equivalent system powered with high-end dual socket X86s—a factor of more than 15X. Our storage system has also set a world record of delivering a staggering 6.55M 4K IOPS to a single storage client! In another example, network cards based on the Fungible DPU deliver well over 25M packets/second while processing complex, stateful rules. Second, the Fungible DPU is unique in that it can dramatically improve the efficiency of node-to-node interactions in a data center while still using a standard Ethernet network. This latter capability, which we call TrueFabric™ enables virtually all data center resources to be disaggregated and pooled for efficiency without burdening CPUs or GPUs. Needless to say, the Fungible DPU is capable of supporting all standards-based network protocols including TCP and ROCEv2, and having them interoperate seamlessly with TrueFabric. The net benefit to using the Fungible DPU in a data center is a 2-3X improvement in CAPEX and a 2X improvement in power consumption for a given level of performance.
For well over a year now, we have been engaged in proofs of concept (POCs) with multiple large customers, including some of the most demanding workloads in the industry coming from more than one customer and more than one vertical. We have received excellent feedback on our products as well as tremendous encouragement in the bold direction we have taken with the Fungible DPU. We are really excited about our early commercial success and are poised for accelerating this success in the coming months.
As we enter this new phase in Fungible’s journey towards a great, category-defining technology company, I have really exciting news to share with you: After an extensive search, we are thrilled to bring Eric Hayes on board as CEO of Fungible. Eric brings great industry experience in scaling organizations and accelerating business momentum with strong customer partnerships. I will continue as Executive Chairman of the Board and take on the new role of Chief Development Officer to ensure that Fungible continues to innovate to provide the most advanced data center products and solutions for our customers.
This is an important step in building a category-defining technology company. With Eric as CEO and myself as CDO, the entire Fungible team will remain relentlessly focused on delivering our vision to customers around the world: to revolutionize the performance, economics, reliability, and security of data centers at all scales and geographies.