How the Analog Processor Can Revolutionize Edge AI

Many companies, from IBM to RAIN Neuromorphic, see the potential, but Mythic is first on the market.

Alberto Romero, Cambrian-AI Analyst, contributed to this article.

Mythic is an artificial intelligence processor company designed to overcome the growing limitations of digital processors. Founded by Mike Henry and Dave Vick, and based in Texas, Austin, and Redwood City, California, Mythic aims to solve the technical and physical bottlenecks that limit current processors through the use of analog computing in a world dominated by digital technology. Mythic wants to prove that, contrary to popular belief, the analog is not a relic of the past, but a promise for the future.

There are two main problems hindering the pace of development of digital devices: the end of Moore’s Law and the von Neumann architecture. For 60 years, we’ve enjoyed ever-increasingly powerful devices as Gordon Moore predicted in 1965, but as we approach the minimum theoretical size of transistors, his well-harnessed law appears to be about to expire. Another known problem is the need in the Von Neumann architecture to transfer data from memory to the processor and back to perform the calculations. This approach is increasingly being replaced by computing in memory (CIM) or near-memory computing methods that dramatically reduce memory bandwidth and response time while increasing performance.

The return of analog computing?

Mythic claims to have built a unique paradigm-shifting solution that promises to address digital limitations while providing improved specifications over best-in-class digital solutions: the Analog Computing Engine (ACE). Historically, analog computers have been replaced by digital computers due to the latter’s lower cost, size, and general purpose nature. However, the current landscape of AI is dominated by deep neural networks (DNNs) that do not require extreme accuracy and, most importantly, the majority of computing goes into a single process: matrix multiplication. An ideal opportunity for analog computation.

Moreover, Mythic exploits the advantages of CIM and data flow architecture to get great early results. They have taken CIM to the extreme by computing directly into flash memory cells. Analog matrix processors take the input as the voltage, the weights are stored as the resistance, and the output is the output current. In addition, the data flow design keeps these processes running in parallel, allowing for very fast and efficient calculations while maintaining high performance. An intelligent blend of analog computing, CIM, and data flow engineering defines Mythic ACE, the company’s key excellence technology.

Mythic’s ACE meets the requirements of advanced AI inference

Mythic technology promises high performance with ultra-low power, ultra-low latency, low cost and a small form factor. The core element is the Analog Matrix Processor (AMP) that features a set of squares, each containing an ACE supplemented with digital elements: SRAM, a SIMD vector module, a NoC router, and a 32-bit RISC-V nano-processor. ACE’s innovative design eliminates the need for DDR DRAM, reducing latency, cost, and power consumption. AMP chips can be scaled, providing support for large or multiple formats. Their first product, the single-chip M1076 AMP (76 AMP segments) can handle multiple endpoint applications and can be scaled up to 4-AMPs or even 16-AMPs on a single PCI Express card, suitable for high performance at the server level. the use.

The hardware is complemented by a software stack that provides a seamless pipeline that goes from graph (ONNX and PyTorch) to AMP-ready package through process optimization (including quantization to analog INT8) and assembly. The Mythic platform also supports a library of ready-to-use DNNs, including object detection/classification (YOLO, ResNet, etc.) and mode estimation models (OpenPose).

The company’s full-stack solution enhances the capabilities of analog processors while preserving features relevant to the digital world. It makes the M1076 AMP a great choice for handling edge inference AI workloads faster and more efficiently — the company claims it provides “better TOPS/W” — than its all-digital counterparts. This, combined with the company’s broad product offering and AI models, makes it well-positioned to target rapidly growing AI-focused markets such as video surveillance, smart home devices, augmented/virtual reality, drones, and robotics.

So far, Mythic appears to have turned an innovative idea into a promising technology to compete for AI heuristics. Now, let’s see the numbers. The company claims that the M1076 AMP operates at up to 25 TOPS operating at around 3W. Compared to similar digital devices, this represents a reduction in power consumption of up to 10x. It can store up to 80 million weights on the chip. The MP10304 Quad-AMP PCIe card can deliver up to 100 TOPS at 25W and store 320Mb weights. When we compare these claims to those of many others, we can’t help but be impressed.

Conclusions

The success of analog AI will depend on achieving high density, high throughput, low latency and high energy efficiency, while delivering accurate predictions. Compared to purely digital applications, analog circuits are inherently noisy, but despite this challenge, the benefits of analog computing are becoming apparent as processors such as the M1076 are able to run larger DNN models featuring higher resolution, higher resolution, or lower latency.

As Mythic continues to improve its hardware and software, we look forward to seeing benchmarks that can demonstrate the platform’s capabilities and energy efficiency. But we’ve already seen enough to be excited about the potential of this unique approach.

.