Fungible today at Hot Chips 2020 announced a new class of microprocessor it calls the Fungible data processing unit (DPU), which won endorsements from Alibaba Cloud, Juniper Networks, and Samsung Electronics.

“We invented what we believe to be the third socket in data centers,” said Pradeep Sindhu, CEO and co-founder of Fungible. Sindhu previously founded Juniper and served as that company’s CEO and CTO.

The technology compliments the CPU and GPU in scale-out data centers by offloading what Fungible calls “data-centric computations” — these include storage, security, and virtualization — onto the DPU. Sindhu says the DPU performs these computations 10-times faster than CPUs. The technology also uses data center resources more efficiently, using less power, and thus providing total cost of ownership savings, he added.

“If you take the best data center from Google or Amazon, we believe that by applying our technology you can improve performance per unit power and space by a factor of two,” Sindhu said. “For enterprise data centers, we believe the improvement can be substantially higher.”

The Fungible DPU platform includes both hardware and software, which disaggregates and composes compute and storage resources on demand. It includes two core pieces: a programmable data-path engine, which executes data-centric computations at high speeds and provides more flexibility compared to general-purpose CPUs. The engine is programmed in C using industry standard tool chains and is designed to execute many data-path computations concurrently.

And the second piece is a network engine that implements the endpoint of Fungible’s proprietary TrueFabric. The high-performance fabric provides deterministic low latency, full cross-section bandwidth, congestion and error control, and high security from hundreds to hundreds-of-thousands of nodes, the company claims.

The TrueFabric protocol is standards compliant and interoperable with TCP/IP over Ethernet. This ensures that the data center leaf-spine network can be built with standard Ethernet switches and standard electro-optics and fiber infrastructure.

Sindhu and Bertrand Serlet founded Fungible in 2015. Prior to Fungible, Serlet founded a storage startup called Upthere, which was acquired by Western Digital, and before that was SVP of software engineering at Apple. At the time, the two set out to solve “two fundamental issues in scale-out data centers,” Sindhu said.

The first is a networking problem. “When one server tries to exchange data with another server that interchange of data is very inefficient,” he said.

The second problem involves data-centric computations. The sheer amount of data is exploding, and general-purpose CPUs are slow and inefficient when it comes to executing functions like SDN, security, storage, and virtualization, Sindhu said. “General purpose CPUs were never designed to execute these computations efficiently, because these computations did not exist when general-purpose CPUs were being designed,” he explained.

Plus, most workloads these days are microservices-based. “So both the style of writing software, as well as the desire to handle very large data sets is what causes the data-centric computations to explode,” Sindhu said.

In June 2019, the startup closed a $200 million Series C funding round led by SoftBank Vision Fund, which brought its total raised to more than $300 million.

And after five years in the making, its DPUs are available to “select design partners.”

They come in two versions. The Fungible F1 DPU is an 800 Gb/s processor designed for high-performance storage, analytics, and security platforms. And the Fungible S1 DPU is a 200 Gb/s processor optimized for host-side use cases including bare metal virtualization, storage initiator, NFV infrastructure/virtual network functions (VNFs) applications, and distributed node security.

Fungible execs won’t name any of these select design partners outright, but its DPU announcement includes accolades from Samsung, Juniper, and Alibaba Cloud executives.

“Bottlenecks in disaggregated infrastructure force users to trade off compression and encryption with storage performance,” said Young Sohn, president and chief strategy officer for Samsung Electronics, in a statement. “The Fungible DPU, which is optimized for data-centric compute, enables line rate storage and security data services, unleashing the full capabilities and blazing fast performance of NVMe SSDs over the network.”

Perhaps not surprisingly, Juniper is working with Fungible “to address large-scale network challenges in data centers and beyond,” said Juniper CEO Rami Rahim in a statement. “Fungible’s TrueFabric technology has demonstrated game-changing performance and latency characteristics that allow efficient disaggregation of all resources at data center scale. In addition, the Fungible DPU’s programmability has the potential to bring data-path agility to networks broadly.”

IDC expects this type of technology — which it calls a function offload co-processor and includes Nvidia GPUs in this category — to become increasingly important in scale-out data centers that want to disaggregate their compute and storage resources.

“When you do everything in software you realize that at some point the CPU is the bottleneck,” said Ashish Nadkarni, group VP of worldwide infrastructure at IDC. “And so it clearly begs for a different hardware architecture. With the DPU, you can sort of bypass the whole CPU dependency by creating you own co-processor clusters. And therefore, your composability now becomes a lot more scalable.”