AMD moves to monolithic 6nm FPGA for edge AI, automotive
AMD has launched its second generation of Versal ‘adaptive SoC’ for edge AI applications, moving back to a monolithic chip rather than using a chiplet design.
The Versal AI Edge Series Gen 2 and Versal Prime Series Gen 2 adaptive SoCs combine pre-processing, AI inference, and postprocessing together in a single 6nm device for end-to-end acceleration of AI-driven embedded systems with ASIL-D levels of functional safety.
Subaru is among first customers to announce plans to deploy AMD Versal AI Edge Series Gen 2 for an ADAS vision system that is purely camera-based and does not need radar or lidar sensors. The chip is also aimed at robotics systems and medical applications.
“You have to think about the whole pipeline and not just the inference,” says Salil Raje, senior vice president and general manager of the Adaptive and Embedded Computing Group at AMD. ”You need adaptive computing as the models are changing and the sensors are changing. It’s more than FPGA, you have to think about the software with heterogeneous platforms,” he tells eeNews Europe.
The devices are aimed at functional safety applications, with up to eight 8x ARM Coretex-A78AE automotive cores and 10 ARM Cortex-R52 real time cores. There are up to a 144 AI engines, where each is a VLIW processor and connects any to any using stream with multiple data types, compute arrays and control tiles with memory tiles.
These AI accelerators can also use the logic fabric to pull additional memory into the AI sub-system, a key requirement for transformer-based models that are often memory bound.
The chip has also been architected for ASIL-D functional safety. The Accelerated Processor Units (APU) with the 78AE cores are in cluster so two that can be used in lockstep, and the R52 cores in the real time processor unit (RPU) are similarly in clusters of two and can be used in lock step. “Having the APU and RPU in lockstep gives us ASIL-D, there are clusters of two each with lock step,” said Raje.
“Subaru is adopting this for the EyeSight AI stereo cameras, no lidar, no radar, no maps,” he said. “They are using Gen 1 for the development and will be using Gen 2, and we give them a 15 year commitment for supply,” he said. A key point is that this is the same engines as the Ryzen AI microprocessors that are used in the data centre to create and train the models, he says, although AMD points out the FPGAs are not used for training.
The EyeSight system is integrated into select Subaru car models to enable advanced safety features, including adaptive cruise control, lane-keep assist, and pre-collision braking.
“Subaru has selected Versal AI Edge Series Gen 2 to deliver the next generation of automotive AI performance and safety for future EyeSight-equipped vehicles,” said Satoshi Katahira, General Manager, Advanced Integration System Department & ADAS Development Department, Engineering Division at Subaru. “Versal AI Edge Gen 2 devices are designed to provide the AI inference performance, ultra-low latency, and functional safety capabilities required to put cutting-edge AI-based safety features in the hands of drivers.”
The Versal Prime devices also include new hard IP for high-throughput video processing, including up to 8K multi-channel workflows,
Designers can get started with AMD Versal AI Edge Series Gen 2 and Versal Prime Series Gen 2 early access documentation and first-generation Versal evaluation kits and design tools available today. AMD expects availability of Versal Series Gen 2 silicon samples in the first half of 2025, followed by evaluation kits and System-on-Modules samples in mid-2025, and production silicon expected in late 2025.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
