Supercharge Your Neural Networks with the Mythic AMP
Table of Contents
- Introduction
- Mythic: A Brief Overview
- Mythic's Unique Technology
- Mythic's Product Line
- Software for Mythic Amp
- Mythic's Workflow for Analog Aware Training
- Binary Generation and Chip Programming
- Mythic's Runtime and System Integration
- Performance Results
- Conclusion
Mythic: Accelerating Deep Neural Networks on Mythic Amp
In today's digital age, the demand for artificial intelligence (AI) and deep learning capabilities is ever-increasing. Companies and researchers are constantly pushing the boundaries of what is possible with AI, and one area that has seen significant advancements is deep neural networks (DNNs). These networks have revolutionized fields such as image recognition, natural language processing, and autonomous driving. However, as the complexity of DNNs continues to grow, so does the need for more powerful and efficient hardware solutions.
1. Introduction
Mythic, a venture-funded startup, has been at the forefront of developing power-efficient inference processing solutions. With a history dating back to 2012, Mythic has focused on leveraging analog computed memory technology to accelerate DNNs. This unique approach combines embedded flash with analog circuits, resulting in a high-density flash memory with exceptional performance at low power levels.
In this article, we will Delve into Mythic's technology, products, and software, providing an in-depth understanding of how they accelerate deep neural networks on the Mythic Amp.
2. Mythic: A Brief Overview
To fully appreciate the significance of Mythic's advancements, it is essential to understand the company's background and vision. With over 110 employees and offices in Austin and Redwood City, Mythic has received over $90 million in venture capital funding. Major investors such as SoftBank, DFJ Lux, and Baylor have recognized the potential of Mythic's analog computed memory technology.
3. Mythic's Unique Technology
The heart of Mythic's innovation lies in its analog computed memory technology, which provides the best of both worlds: high-density flash memory and high-performance analog compute engines. This unique combination eliminates the need for external DRAM and delivers a deterministic execution model through its data flow architecture. These features result in an exceptional combination of latency, power efficiency, and weight capacity in a single chip.
4. Mythic's Product Line
Mythic has recently introduced its flagship product, the M1108, which is the industry's first analog matrix processor (AMP). This powerful processor boasts a data flow architecture with 112 tiles, of which 108 are equipped with analog compute engines (ACEs). Each compute tile includes an analog engine, a digital CMOS vector engine, a 32-bit RISC-5 nanoprocessor, a network-on-chip router, and substantial local SRAM. With support for N4 into 8 and N16 computation, the M1108 can handle well over 100 million weights and run multiple networks entirely on chip.
5. Software for Mythic Amp
To fully harness the potential of the Mythic Amp, Mythic has developed a comprehensive software ecosystem. The software is designed to provide an easy and familiar flow, allowing users to achieve high performance without the need for HAND coding or optimization. Mythic aims to support a broad range of DNNs, network types, and training frameworks, ensuring exceptional performance and image processing accuracy. The software's flexibility also enables the efficient utilization of the Mythic Ace tiles and digital processing resources.
6. Mythic's Workflow for Analog Aware Training
Mythic's software workflow simplifies the process of analog aware training, making it as easy as quantization for traditional processors. The workflow includes transforming the network to an analog aware network through quantization, optimization, and retraining that considers the strengths of Mythic's analog compute engines. The mythic graph compiler optimizes memory usage and enables maximum performance for image recognition tasks.
7. Binary Generation and Chip Programming
Once the analog aware network is ready, the graph compiler generates the binary code for the Mythic processor. This code can be compiled into firmware and programmed into the Mythic Amp. The binary generation stage also includes the integration of the firmware library, the RISC-5 compiler, and static routines. Additionally, the compiler can generate different outputs depending on the target, allowing for various host platforms and system configurations.
8. Mythic's Runtime and System Integration
Mythic provides a lightweight host runtime that manages input data, controls the Mythic Amp, and returns the results to the host system. The runtime supports common operating systems and is easily portable to new ones. It includes a library of low-level utilities for customer use and debugging and provides lightweight and portable drivers for managing multiple Mythic Amps in Parallel.
9. Performance Results
The performance of the Mythic Amp is truly impressive. Preliminary results Show low latency and excellent accuracy, rivaling that of digital solutions. For example, in running ResNet50, the Mythic Amp achieves nearly 900 frames per Second at a resolution of 224x224. The performance for other networks, such as YOLO, is also promising, with the potential to reach up to 60 frames per second at a resolution of 416x416.
10. Conclusion
Mythic's analog computed memory technology and the Mythic Amp represent a significant leap forward in deep neural network acceleration. With their unique solution, Mythic provides a powerful, efficient, and cost-effective alternative to traditional digital processors. The ease of use and flexibility of Mythic's software ecosystem further enhance the capabilities of the Mythic Amp, making it an attractive choice for a wide range of AI applications.
As the era of analog compute dawns, Mythic is poised to drive innovation and revolutionize the field of deep learning. With their ongoing advancements in hardware and software, Mythic is enabling organizations to unlock the full potential of neural networks and propel artificial intelligence into the future.