An announcement yesterday about a $200M investment in a data center silicon startup in Silicon Valley was a momentous revival of the glorious past. This has not happened in a while, in such a mammoth scale. Even on a smaller scale, the focus lately has been on silicon related to machine learning and inferencing. Not on a piece of silicon that feels and smells like a networking chip as it directs traffic efficiently to multiple compute elements most commonly found in modern data center servers. There was also a lot of intrigue and even surprise yesterday - for many of us in the data center infrastructure silicon industry. The announcement that Fungible raised $200M of series C funding with SoftBank as a lead investor raised many eyebrows. Many of my colleagues in the industry asked how and why. Some asked, what is a DPU? Others wondered why so much money is needed so early in the life of a startup.
Fungible is led by an industry visionary and founder of Juniper Networks, Pradeep Sindhu. I have followed his latest initiatives in the data center. Most noteworthy were the launch of the QFX data center switch series and QFabric architectures at Juniper, the acquisition of Contrail and foray into SDN by Juniper, and most recently the introduction of the Data Processing Unit or DPU. The QFX and QFabric were a repositioning of the then well-understood leaf and spine switch fabric and topologies that were becoming very prevalent in hyperscale data centers. Contrail was an alternative virtual router and MPLS-based SDN solution as opposed to the significant movement with virtual switch-based SDN solutions at hyperscale data centers, and also introduced by Nicira to the general market using the Open vSwitch (OVS) and the NSX Controller. Nicira was acquired by VMware and the acquisition of Contrail by Juniper followed soon after.
Fast-forward to taking a closer look at the DPU announced recently. To me, Sindhu is following a similar approach - riding on a prevalent trend with a slightly different take, or must I say, “repositioning."
If one looks carefully at the DPU in the figure (reproduced from yesterday’s announcement), it is an element that is receiving traffic (and data in the form of packets) from the network and directing them to other compute (CPU, GPU, FPGA in the figure) and storage (SSD in the figure) elements that typically reside inside a modern data center server. Traditionally, the Network Interface Card (NIC) in the server would receive the traffic from the network and rely on the host CPU (in almost all cases an x86 CPU) to direct the traffic to those compute and storage elements. The Fungible DPU now takes the role of the NIC and the director of traffic offloading that task from the host or x86 CPU.
I have been in the world of NICs for almost 10 years, and in the world of SmartNICs and heterogeneous computing for the last 5 years. To me, the DPU looks like SmartNIC silicon with accelerators that assist the directing of network packets (and the data they carry) efficiently to the compute elements (including technologies like GPUDirect for GPUs), and in case of the SSDs, also assist with certain compute tasks (encryption, compression, deduplication, erasure coding etc.). SmartNIC adoption is skyrocketing, led initially by the hyperscaler data center operators. Heterogeneous computing and domain specific accelerators are helping bring significant efficiencies in data center infrastructure. These are significant waves to ride on and the DPU is doing exactly that, albeit with a different name.
According to the announcement yesterday, Fungible was founded four years ago, has about 200 employees and has raised more than $300 million in total funding. That’s a lot of money. Silicon development and manufacturing, especially in smaller process nodes, has become very expensive. Back in Oct of 2018, Sidhu was quoted supporting the MIPS CPU architecture, “… MIPS technology is a key component of the Data Processing Unit we are developing as a fundamental new building block of next generation data centers…” SoftBank owns ARM Holdings and is a proponent of Arm-based CPUs, so I suspect Fungible will have to change course and adopt Arm-based DPUs – incurring significant future investments in silicon. The use of chiplets may be an effective way to control such skyrocketing costs.
The news is definitely intriguing and groundbreaking. It highlights the important role that a piece of silicon plays – the one that receives packets from and transmits packets to the network and then directs them efficiently to compute and storage elements that reside inside a modern data center server. Arguing whether that component is a DPU or SmartNIC silicon is beside the point. The significant investment amount reinforces the trend toward heterogeneous computing. The announcement also highlights the burgeoning costs of silicon development and manufacturing, and the need to control them with innovations around chiplets, another trend that the hyperscalers and large silicon companies have already adopted.