Automotive ethernet prompts Marvell to splash $452mn on Aquantia

Marvell has snapped up an ethernet-focused rival, Aquantia, for $452mn – a 39% premium on its closing share price. While there is considerable opportunity for Marvell to use Aquantia’s technologies in conventional data center applications, specifically its high-bandwidth transceivers, Marvell has pointed to the growing demand for in-vehicle networking in the automotive sector, where Aquantia has a few semiconductor designs that cater to autonomous vehicles.

Though we loathe the term ‘data centers on wheels,’ it would make sense for cars to begin using some of the data center technologies available in their own automotive applications. While we probably aren’t going to be swapping the spare tires for 2U racks, or CAN-bus for InfiniBand, early automotive experiments with ethernet have shown that the networking technology has a bright future in vehicles.

Aquantia has worked with Nvidia on its Pegasus Drive AGX platform, which Nvidia hopes is going to be the sensor-fusion hub for self-driving cars, and is also a member of the Networking for Autonomous Vehicles (NAV) Alliance. Nvidia also picked Marvell’s automotive ethernet switch (88Q5050) to be integrated into the Drive Pegasus platform, meaning that by buying Aquantia, Marvell seems to be snapping up a rival in what could be the early stages of a booming market.

This view would require seeing Aquantia as a force to be reckoned with, which doesn’t seem to be the case based on its most recent quarterly results – posting a $13mn loss on revenues of $17mn. Aquantia’s IPO, back in 2017, closed below its valuation range, at $9, rather than the $10-12 range, achieving a valuation of $290mn. Perhaps it was getting ready to turn a corner, but either way, Marvell has snapped up a future competitor – for a lot less than what it paid Intel-botherer Cavium, where Marvell paid $6bn to get a hold of the ARM-based CPU slinger.

Notably, though, Marvell and Aquantia’s customer, Nvidia, has also made an acquisition in this realm. It bought Mellanox, snatching it from Intel in a $6.9bn deal confirmed in March. Ostensibly, that was to advance Nvidia’s data center portfolio, but as both Marvell and Aquantia discovered, those data center designs can be translated to better suit the automotive sector.

The wiring looms in modern cars are incredibly complicated, but also incredibly fragile. Bundles of wires, snaking through bodywork and holes in firewalls to connect the different controller boards to their respective components. Just one over-tightened clip or a poorly crimped cable can lead to many hours of pained diagnostics, as a technician tries to find the source of the loss of connection between a switch or sensor and its controller, in the serpentine loom.

Adding components to a car, as often happens in performance-focused upgrades or upkeep on older vehicles that might necessitate an engine or transmission swap, often requires either extensive modifications to the wiring loom, or using the loom from the donor engine inside the old car.

Now, it would be an oversimplification, but feasibly, ethernet might allow for just a few cables to be used to connect all the components in a car. A connection from the cockpit, from the engine, from the drive train, and from the sensors, all linking to the centralized computer or segregated controller systems.

In that world, automakers would not have to have multiple different wiring loom configurations for each of their car models, which might also vary by the geographic market they are intended to be sold in. Instead, the automaker might be able to standardize on the hardware, use far fewer interchangeable components, and rely on using software to accommodate these different configurations, rather than actually physical differences.

Computer units linked by an ethernet run (perhaps two for redundancy) would be a lot easier than trying to graft such capabilities onto things like this or this. A very simplified comparison might see a few Intel NUC boxes linked into a network via Ethernet, with the NUCs each controlling a distinct function within the vehicle. In that model, you just have to replace the NUC, or more likely, change the software running on the NUC. You wouldn’t have to change the ethernet cabling to match a particular geography or trim-level.

Ethernet would also provide the bandwidth needed to transport all the sensor data around the vehicle, much of which is going to be very taxing, such as stereo camera feeds, LiDAR streams, and radar point clouds. This data needs to get to the processors governing the safety features that enable emergency braking, as well as things like self-parking, and in time, fully autonomous vehicles.

What’s more, Ethernet would bring with it the ability to use much of the IP-based technologies that have emerged in IT, which should provide a pretty smooth way to introduce security features like network firewalls, hypervisors and/or containers, and traffic monitoring. These would ensure that the vehicle’s internal traffic could be monitored, to ensure that nothing malicious has managed to get onboard and that specific components are not overstepping their bounds. It would also facilitate improved security in interactions between the vehicle and external systems too.