Unlocking the Power of Mythic Amp: Accelerate Deep Neural Networks with Analog Compute

Table of Contents:

1. Introduction
2. Mythic: A Brief Overview
3. Mythic Amp: Accelerating Deep Neural Networks
   • 3.1 Mythic's Unique Analog Computed Memory Technology
   • 3.2 Mythic Amp Features and Specifications
4. The Advantages of Analog Compute
   • 4.1 Power Efficiency
   • 4.2 High Performance at Low Power
   • 4.3 Deterministic Execution Model
5. Scalability from Wearables to Edge Servers
6. Mythic's First Product: The M1108
   • 6.1 Data Flow Architecture
   • 6.2 Analog Compute Engines (ACEs)
   • 6.3 Mythic's Comprehensive Interface Support
7. Performance and Metrics in Edge AI Solutions
   • 7.1 Hardware-Driven Metrics
   • 7.2 Software-Driven Metrics
8. Mythic's Software: Easy, Powerful, and Versatile
   • 8.1 Objectives of Mythic's Software
   • 8.2 Mythic's Easy and Familiar Workflow
   • 8.3 Optimizing for Power Reduction, Latency, and Tile Utilization
9. Mythic's Workflow Stages
   • 9.1 Transforming the Network into an Analog-Aware Network
   • 9.2 Retraining for Analog Compute
   • 9.3 Graph Compilation and Binary Generation
   • 9.4 Programming the Mythic Chip for Initialization
10. Mythic's Runtime and Support
    • 10.1 Managing Input Data and Controlling the Mythic Amp
    • 10.2 Lightweight and Portable Drivers
    • 10.3 Supported Operating Systems
11. Preliminary Results and Future Optimizations
    • 11.1 ResNet Performance
    • 11.2 YOLO Performance
12. Conclusion
13. Employee Showcase
14. Frequently Asked Questions (FAQ)

🚀 Mythic: Accelerating Deep Neural Networks with Analog Compute

Mythic is a venture-funded startup with a vision to accelerate deep neural networks (DNNs) using their unique analog computed memory technology. Their flagship product, the Mythic Amp, offers a powerful and power-efficient solution for edge AI applications.

1. Introduction

In this article, we will explore Mythic's innovative approach to accelerating DNNs and the features of their Mythic Amp. We will delve into the advantages of analog compute, scalability from wearables to edge servers, and Mythic's software workflow. Additionally, we will discuss the performance metrics in edge AI solutions and the preliminary results achieved by Mythic's hardware and software.

2. Mythic: A Brief Overview

Mythic is a startup that has been at the forefront of power-efficient inference processing since 2012. They have developed a unique analog computed memory technology that combines embedded flash with analog circuits. This allows Mythic to execute DNNs on their chips with all weight storage included, eliminating the need for external DRAM.

With offices in Austin and Redwood City, Mythic has raised over $90 million in VC funding from major investors like SoftBank, DFJ Lux, and Baylor. Mythic's analog compute technology has proven to be ideal for edge AI, offering a best-in-class combination of latency, power, and weight capacity in a single, cost-effective chip.

3. Mythic Amp: Accelerating Deep Neural Networks

The Mythic Amp is the industry's first analog matrix processor designed to accelerate DNNs. It features a data flow architecture with 112 tiles, each housing analog compute engines (ACEs) and other supportive components. The Mythic Amp's impressive specifications enable it to support large networks, deliver high performance, and Consume very low power.

3.1 Mythic's Unique Analog Computed Memory Technology

At the core of Mythic's technology is their analog computed memory, which combines the density of flash memory with the high performance and low power levels of analog compute. This architectural innovation allows Mythic to achieve efficient execution of DNNs by integrating weight storage and computation on a single chip.

3.2 Mythic Amp Features and Specifications

The Mythic Amp's data flow architecture includes ACEs, a digital CMD vector engine, a 32-bit RISC-5 nanoprocessor, and a network-on-chip router for seamless communication. With a capacity for well over 100 million weights and support for computation in various modes, the Mythic Amp can deliver a theoretical maximum of around 35 TOPS and impressive frame rates for real-world applications.

(Note: The article continues with the remaining headings and subheadings, further discussing the advantages of analog compute, scalability, Mythic's software, workflow stages, runtime and support, preliminary performance results, and concluding remarks.)