Mythic

Definition

Mythic is a semiconductor company that designs analog computing chips for AI inference, performing the core multiply-and-accumulate math of neural networks directly within flash memory cells using analog electrical signals rather than the fully digital circuits conventional processors rely on. This in-memory analog compute approach is intended to reduce the power consumption and data movement overhead of running trained AI models at the edge, on devices with tight power and cost budgets rather than in a data center.

Overview

Mythic was founded around a contrarian bet in AI chip design: instead of pursuing ever more digital compute throughput, as most AI processors do, the company set out to perform neural network math using analog electrical properties directly inside flash memory cells, the same type of non-volatile memory used in solid-state drives. The motivation was that a large share of the energy and time cost in running a neural network on digital hardware comes not from the arithmetic itself but from repeatedly moving weight values between memory and compute units, an overhead that in-memory analog computation can largely sidestep for the specific workload of inference on an already-trained model. Technically, Mythic's chips store a trained neural network's weights as precise analog charge levels within flash memory cells, then use the physical properties of current flowing through those cells to perform the multiply-and-accumulate operations that dominate neural network inference, with results converted back to digital signals for further processing. Because the computation happens where the weights are stored rather than requiring them to be fetched and moved to a separate compute unit, this in-memory analog approach can substantially cut the power draw associated with data movement, at the cost of the precision and flexibility that fully digital compute provides, and it is generally suited to inference on models already trained elsewhere rather than to training new models from scratch. Among AI hardware companies, Mythic occupies a narrow, specialized niche distinct from data-center-focused chipmakers like Cerebras, SambaNova, and Graphcore, which target large-scale training and inference with digital architectures. Mythic instead competes in the edge AI inference space, alongside other low-power AI accelerator approaches, where power efficiency and cost per unit matter more than raw throughput or training capability. In practice, Mythic's analog compute chips have been positioned for embedded and edge applications such as smart cameras, industrial sensors, robotics, and other devices that need to run AI inference locally under strict power constraints, avoiding the latency and connectivity dependence of sending data to a cloud server for processing. This makes the technology relevant to product categories where battery life or thermal budget rules out a conventional digital AI accelerator. The analog approach carries inherent trade-offs. Analog computation is generally less precise than digital arithmetic and can be more sensitive to manufacturing variation and environmental conditions such as temperature, which constrains the range of neural network types and precision requirements it suits well, and because it is inference-only and edge-focused, it does not compete with digital accelerators or GPUs for large-scale model training, making it a complement to, rather than a replacement for, data-center AI hardware.

Key Features

• Analog in-memory compute using flash memory cells for neural network math
• Designed specifically for AI inference rather than model training
• Reduced data movement overhead compared to digital compute architectures
• Targeted at low-power edge and embedded AI applications
• Trade-off of analog precision limits versus digital compute flexibility
• Suited to devices with strict power and cost constraints

Use Cases

Alternatives

Frequently Asked Questions

Related in AI Tools