Processing In Memory: The Mythic Advantage
Deep neural networks demand massive parallel compute capabilities that weigh heavily on conventional local AI – but not on Mythic’s platform, which performs hybrid digital/analog calculation inside flash arrays. This entirely new approach performs the inference step of deep neural networks inside the memory array that stores the processing weights long term – bringing huge advantages in performance, power life and accuracy.
Desktop GPU Capabilities in a Shirt Button-sized Device
Mythic’s platform delivers the power of desktop GPU in a button-sized chip that supports multi-million weight neural networks. Using rapid scaling technology, it delivers massive parallel computing when intelligence is needed then vanishes when not. The result is near-weightlessness in terms of device constraints – but all the power you need to support deep neural networks.
Features and Benefits
- Supports all popular styles of deep neural networks, including convolutional, recurrent, and fully connected.
- Zero performance issues: The weight storage also handles the processing, so no matter the size of the network, processing capabilities will always be sufficient.
- The massive parallel processing of this approach means unmatched benefits in terms of low latency and high throughput.
- 50x lower power: Co-located processor/storage and massively parallel analog arithmetic deliver huge energy efficiencies compared to all-digital platforms.
- Absolute accuracy: Mythic’s techniques ensure negligible accuracy loss during neural network inference compared to an all-digital system.
- A software development environment that interfaces directly with machine learning packages such as TensorFlow. We make it trivial to push any deep neural network ensemble onto our platform.