As appeared on Embedded Computing Design:

GTI Releases AI Accelerator ASIC with Integrated MRAM


Gyrfalcon Technology Inc. (GTI) has released the Lightspeeur 2802M AI Accelerator, a 22 nm production-ready ASIC with 40 MB embedded MRAM. The 2802M is the first production-ready accelerator with embedded MRAM, the company says, and can support large or multiple AI algorithms for image classification, voice identification, voice commands, facial recognition, or pattern recognition on a single device.

The 2802M is the first product in GTI’s MRAM-based product roadmap. The device combines an MRAM engine with the patented, convolutional neural network-based (CNN-based) Matrix Processing Engine (MPE) and AI Processing in Memory (APiM) technologies to mimic the logic and memory of a human brain. This architecture helps achieve high tera operations per second (TOPS) performance at low power consumption because the overhead of managing discrete memory components is avoided.

The ASICs are delivered with access to an SDK and the company’s PLAI Builder, which simplifies AI model creation for lay developers. AI models created in Caffe, TensorFlow, and other frameworks can also be supported, as can neural networks like ImageNet, MobileNet, ResNet, SDD, VGG, etc.

The 2802M was delivered to TSMC in June 2018. GTI will be showcasing its solutions at the Consumer Electronics Show (CES) in Las Vegas in January.

For more information, visit