Last week, Mythic announced a breakthrough with compute-in-memory technology based on a 40 nm process with what the company claims to be the industry’s first Analog Matrix Processor. The M1108 AMP AI accelerator chip targets high-end edge AI applications including smart home, AR/VR, drones, and is said to set a benchmark in the industry for high performance and low power in a single cost-effective device, also available in M.2 and PCIe form factors.
The M1108 comes with an array of flash cells, ADCs, a 32-bit RISC-V nano-processor, a SIMD vector engine, SRAM, and a high-throughput Network-on-Chip (NOC) router. With 108 AMP tiles, the M1108 provides up to 35 Trillion-Operations-per-Second (TOPS) enabling ResNet-50 at up to 870 fps. This enables a power-efficient execution of complex AI models such as ResNet-50, YOLOv3, and OpenPose Body25. The industry leader NVIDIA also has a similar AI accelerator chip NVIDIA Xavier AGX which delivers up to 32 TOPS.
A wide variety of host processors are supported, including Intel x86, NXP iMX8, NVIDIA Jetson, and Qualcomm RB5. The power consumption of M1108 while running AI models at peak throughput is around 4W.
The M1108 AI accelerator chip comes with powerful pre-qualified models for the high-end edge AI use cases. Available pre-qualified models in development include object detector and classifier, human pose estimator, image segmentation, just to name a few.
The M.2 card (22mm x 80mm) comes with a compact form-factor which makes it easy for integration into many different systems. This card is ideal for processing deep neural network (DNN) models for applications mentioned above. It can execute multiple DNNs concurrently. Also has 4-lane PCIe 2.1 for up to 2GB/s bandwidth and no external DRAM. The plan is to support Ubuntu, NVIDIA L4T, and Windows (future release).
This PCIe evaluation card (156mm x 121mm) will allow us to easily evaluate Mythic’s high performance, yet power-efficient AI inference solution for edge devices and servers. The AI workflow has support for standard frameworks, including PyTorch, TensorFlow 2.0, and Caffe.
Without installing software or drivers, the evaluation system (4.25″ L 4.25″ W 1.75″ H) is the most convenient way to evaluate the performance.
The M1108 AMP will be available in both PCIe M.2 and PCIe card form-factors, and M1108 mini PC evaluation kits are available on request.
Source: All the figures used are from the Mythic website.
Abhishek Jadhav is an engineering student, RISC-V Ambassador, freelance tech writer, and leader of the Open Hardware Developer Community.
|Support CNX Software - Donate via PayPal or cryptocurrencies, become a Patron on Patreon, or buy review samples|