Dream Chip Technologies (DCT), a 25-year-old German company, announced record power efficiency for its ADAS system-on-chip (SoC) targeted at automotive computer vision applications. Fabricated on GlobalFoundries’ (GF) 22FDX semiconductor process at the foundry’s Fab 1 facility in Dresden, Germany, the SoC was developed as part of the European Commission’s ENIAC THINGS2DO ADAS reference development platform program. Up to 40 partners in the program across Europe are cooperating to champion the FD-SOI design ecosystem.
The objective here was to create a camera-based ADAS reference platform that benefits automotive Tier 1 companies. The SoC was created in close cooperation with Arm, ArterisIP, Cadence, GF, and INVECAS, which provided IP and end-to-end ASIC design solutions as a part of GF’s FDXcelerator Partner Program.
GF claims that 22FDX delivers FinFET-like performance and energy efficiency at a cost comparable to 28-nm planar technologies. According to the foundry, FDX technologies represents an alternative path for applications that can’t accept the cost and complexity of FinFETs. 22FDX is said to deliver a 20% smaller die size and 10% fewer masks than 28 nm, as well as nearly 50% fewer immersion lithography layers than foundry FinFET.
DCT's ADAS SoC for automotive computer vision.
As one of two current leading-edge CMOS technology options (along with the three-dimensional FinFET), for designs where performance and power are both important and overall processing efficiency and battery life are key, fully-depleted silicon-on-insulator (FD-SOI) provides significant low-power, low-cost, and processing efficiency advantages, as often found in mainstream mobile application processors, wireless networking, Internet of Things (IoT), wearables and smart sensors. FinFET, on the other hand, tends to offer advantages for the higher-performance designs that have significant digital content. These designs often show up in high-performance computing, server, or premium smartphone applications.
FD-SOI was originally developed by STMicroelectronics at 28 nm. It was subsequently licensed to Samsung and GF, who took it to 22 nm and now have a family of processes under the umbrella 22FDX.
The SoC is said to deliver high-performance image acquisition and processing capabilities and supports AI/neural-network (NN) vision operation with a total of 1 TOPS (teraoperations per second) at 500 MHz on four parallel engines. With all functions including quad-core ARM Cortex-A53, Tensilica DSPs, and LPDDR4-Interfaces activated, the SoC shows single-digit power dissipation without the need for forced cooling, which is of significant importance for embedding in automotive environments.
The SoC incorporates DCT’s image signal-processing pipeline, working in conjunction with Cadence Tensilica Vision P6 DSPs and a quad-core cluster of ARM Cortex-A53 processors. In addition, a lock-step pair of ARM Cortex-R5 processors provides functional safety; the SoC is interconnected with an ArterisIP FlexNoC network-on-chip. Memory bandwidth is provided by a dual-channel LPDDR4-interface from INVECAS. Two DDR-memory chips and the SoC are mounted together on a chip carrier, so that the module provides 4 GB in total system memory.
The module is the centerpiece of a new ADAS platform from DCT targeted at automotive applications with a need for cost, performance, and low power for embedding into the car without the need for forced cooling, such as via fan or liquid. It’s intended to take over the central image-recognition and manipulation tasks, based on camera capture. Due to its tiny power footprint, it will likely be integrated with the camera module.
Of particular importance is the SoC’s new and reduced power footprint—most competing solutions need active cooling. Implementation of DCT’s SoC on the GF 22FDX platform demonstrated single-digit (1.0) watt and cooling targets for designers managing power dissipation. If needed, DCT said the SoC could increase performance to 2 TOPS at 1.0 GHz by applying GF’s forward body-bias capabilities and other optimization techniques.
"Building the best power-efficiency and machine-learning performances with a fully integrated SoC chip will pave the way for self-driving cars and accelerate ADAS adoption,” says Sanjay Charagulla, Senior Director-Vertical Market Segments, GlobalFoundries.