Multicore Micro Pairs Up with CNN ML Accelerator

Oct. 21, 2020

The MAX78000 from Maxim Integrated adds a CNN machine-learning accelerator to a dual-core microcontroller.

Maxim Integrated is a well-known electronics supplier, delivering everything from micros to power-management solutions. Artificial intelligence (AI) and machine learning (ML) were not part of the mix—until now. Its MAX78000 pairs a mixed multicore CPU with a convolutional-neural-network (CNN) machine-learning accelerator (Fig. 1).

1. The MAX78000 contains Cortex-M4F and RISC-V CPU cores plus a CNN accelerator to efficiently handle machine-learning models.

The MAX78000 has a number of out-of-the-ordinary features, including a unique ID. Another interesting feature is the mixed pair of CPU cores: a 100-MHz Cortex-M4F and a 60-MHz RISC-V. These share main memory. It’s probably the first time these cores have been combined in a single chip, but it may not be the last. The Cortex-M4F is the main processor, while the RISC-V handles I/O chores, such as the care and feeding of the other unique item, the CNN accelerator.

The two cores have access to all of the peripherals. This allows developers to choose the platform that might work best for their computation or peripheral requirements. Most developers will be working in C or C++, so the actual processing core usually will not be an issue. It does make for an interesting build environment.

The CNN accelerator, designed by Maxim Integrated, is optimized for high efficiency and low power. ML models can handle voice and video applications in real time. The idea is to provide these types of services without having to go to the cloud.

The CNN accelerator utilizes 64 processors. These can handle a 1023- × 1023-input per channel in streaming mode or 181 × 181 per channel when the data is preloaded. It supports 1-, 2-, 4-, and 8-bit weights with 432 kB of storage reserved for this use, which can translate to 3.5 million weights. It supports up to 64 layers in a model if pooling is applied at every other layer; otherwise, the limit is 32 layers.

The processer complex and peripherals are on par with Maxim’s other microcontrollers. There are four very-low-power comparators and an eight-channel, 10-bit analog-to-digital converter (ADC). The chip can handle a digital camera as well as acting as a 1-Wire master. Maxim sells a host of 1-Wire peripherals.

The chip features 512 kB of flash memory for applications and 128 kB of SRAM. On-chip encryption and secure boot support is included as well.

The $168 MAX78000EVKit development kit for the MAX78000 supports audio and video development (Fig. 2). A smaller module will also be available as a low-cost evaluation platform. Both are supported by a host of third-party tool vendors. The CNN accelerators are supported by TensorFlow and PyTorch tools.

2. The developer kit (left) offers access to the full functionality of the MAX78000, while the compact evaluation module (center) provides a less expensive way to check out the chip. Maxim has garnered wide third-party support (right) for the platform.

The chip targets applications such as battery-operated environments, where low power is a requirement. The CNN accelerator is designed to deliver low-latency solutions while reducing the power consumption, especially when compared to running algorithms on the Cortex-M4F or the RISC-V core.