Mythic claims M1108 AMP is industry’s first AI analogue matrix processor
Patented analogue compute-in-memory technology is used by Mythic for the M1108 Analog Matrix Processor (AMP). It addresses edge AI requirements of high performance and low power in a single, cost-effective device, says the company.
According to Mythic, this is the first time and analogue compute device has achieved best in class performance and power with accuracy comparable to digital devices. This means that the M1108 AMP can be deployed at the edge in a broad range of applications and markets, including smart home, AR/VR, drones, video surveillance, smart city, and automation on the factory floor. Mythic claims that the technology, places the M1108 AMP “at the forefront of a major new trend in AI processing”.
Running AI models at the edge rather than in the cloud means that designs can be simpler and that privacy can be enhanced, says Mythic. To deploy AI at the edge means that millions of devices need to be deployed, so they must be low in cost and have a small form factor, and must also exhibit low latency, high performance, and low power. Digital inference solutions based on SoCs, Tensor processing units (TPUs), CPUs, GPUs and FPGAs can meet some of these challenges, but are limited due to memory, clock speeds, and process technology, reports Mythic. These trade-offs result in only processors capable of high-performance analogue compute can address them all. Mythic says the M1108 Mythic AMP is such a device.
Mike Henry, co-founder, CEO and chairman of Mythic believes the introduction brings the industry to “a significant inflection point”. “We are delivering technology that was previously thought to be impossible. Our Analog Compute Engine (ACE) eliminates the memory bottlenecks that plague digital solutions by efficiently performing matrix multiplication directly inside the flash memory array itself,” he explained.
The M1108 integrates 108 AMP tiles, each with a Mythic ACE featuring an array of flash cells and ADCs, a 32-bit RISC-V nano-processor, a SIMD vector engine, SRAM, and a high-throughput Network-on-Chip (NoC) router. Four control tiles also provide a high-bandwidth PCIe 2.0 interface to a system host processor. With 108 AMP tiles, the M1108 delivers up to 35 trillion operations per second (TOPS) enabling power-efficient execution of complex AI models such as ResNet-50, YOLOv3, and OpenPose Body25 on a single-chip with extremely low latency, says Mythic.
Typical power consumption while running complex AI models at peak throughput is approximately 4.0W. The use of mature 40nm technology and no need for external DRAM or SRAM, also means that the M1108 has a cost advantage of up to a factor of 10, over comparable digital architectures, claims Mythic.
The M1108 AMP will be available in both PCIe M.2 and PCIe card form-factors. M1108 PCIe evaluation kits are available on request.