TOPS priorities shift for ‘Datacenter on Wheels’

By Junko Yoshida for the Ojo & Yoshida Report – In the ADAS/AV market, the race to a trillion operations per second (TOPS) continues to escalate. At CES 2022. Ambarella, Mobileye and NXP/Hailo shared their diverging approaches to the TOPS race.

What’s at stake?
As the activity around CES 2022 makes clear, tech suppliers are hurtling toward more computing power, AI acceleration, and higher-resolution sensors for automotive platforms. But as compute and sensor capabilities grow exponentially, so does the pressure on auto OEMs to pick the right building blocks and implement them to optimal effect in their next-generation cars.

At the 2011 Tokyo Auto Show, Toyota unveiled a concept car it called a “smartphone on wheels.” Time-hop a decade or so to CES 2022, where connected, software-defined, AI compute-intensive vehicles featuring loads of sensors and data processing power have a new descriptor: “datacenters on wheels.”

Like it or not, the name will probably stick, given that more sensors (to generate vastly more data) and more compute power (to process and fuse data for decision-making and path planning) are prerequisites for any next-generation vehicle architecture.

In the ADAS/AV market, the race to a trillion operations per second (TOPS) continues to escalate. Leading central-compute SoC suppliers Nvidia, Intel/Mobileye, and Qualcomm are applying their datacenter expertise to next-generation vehicles. And as vehicles acquire more compute power, machine-learning–driven applications are proliferating.

Reflecting those trends at CES this year are Ambarella, Mobileye, and NXP Semiconductors.

Ambarella, known for vision processors, made a surprise move by jumping into the TOPS fray with a new automotive SoC family, CV-3, capable of 500 equivalent TOPS (eTOPS) of AI compute.

Mobileye, on the other hand, is taking a less-is-more approach. It unveiled the EyeQ Ultra, running at 176 TOPS, which the company said will meet the cost, performance, and power consumption requirements for consumer autonomous vehicles (AVs). Mobileye touted its “unique perspective into the exact requirements for the self-driving system,” as the company has been engaged in vertical AV development ranging from hardware and software to mapping and service models.

NXP has tipped its strategy to take vehicle machine-learning resources beyond ADAS-level perception. Teaming up with Hailo, an Israeli-based AI accelerator startup, NXP will run new machine-learning applications on vehicle gateway processors (such as NXP’s S32G family, Layerscape) or the central-compute parts that interact with all data throughout the vehicle. The machine-learning apps range from distracted-driver, occupancy, and road- condition monitoring to safety, insurance applications, and vehicle health measures.

Ambarella joins TOPS raceWith the newly announced CV-3 High SoC, the highest-performance version of the CV processor to date, “it looks like Ambarella wants to be a big TOPS player now,” said Phil Amsrud, senior automotive analyst at IHS Markit. This is a change in direction for the vision processor company, which earlier had dismissed the importance of TOPS performance.

CV-3 High block diagarm – Courtesy of Ambarella, 2022

CV-3 High offers a neural vector processor with 500-eTOPS (8-bit) AI compute and 1,000-eTOPS (4-bit) AI performance… Full article