Catalyzing the Edge-AI Revolution with In-Memory Computing (IMC)


03 June 2024




Initial release

Key Points

  • IMC significantly reduces data movement between memory and processor, directly enhancing speed and energy efficiency for data-intensive applications like AI or Digital Signal Processing (DSP).
  • The IMC market is project to see rapid and sustained growth over the next decade, with market research surveys indicating a strong shift towards these technologies in commercial sectors.
  • The lack of a mature software ecosystem, and hardware integration complexity, pose significant barriers to the wider adoption of IMC technologies.

Edge device manufacturers are driven to constantly enhance their product offerings in order to meet the demands of their customers and remain competitive. In recent years, this has meant balancing between the need for incorporating increasingly performant data-centered algorithms, such as AI and DSP, and a pressing demand for user privacy and fast throughputs, both requiring on-device processing. Reconciling these demands is challenging in resource-constrained edge devices, as modern computer performance is bottlenecked by the need to shuffle data between the device’s memory and processor, which is also a very energy intensive operation. Thus, overcoming this situation requires a paradigm shift.

In-memory computing (IMC) offers a groundbreaking solution to these challenges by integrating processing within memory arrays. This eliminates the traditional data transmission bottleneck by reducing data movement between the memory and processor, freeing up both the data transmission bus and processor, lowering power consumption and increasing throughput by up to several orders of magnitude. This enhanced performance is pivotal for enabling real-time, on-device data-centric applications such as AI and DSP, paving the way for the next generation of edge devices, from autonomous cars and robots to wearables and customizable sensors.

Adopting and integrating IMC into existing products is, however, challenging and highly disruptive to established hardware and software ecosystems. Besides necessitating hardware redesigns to tackle the complex data control and coordination problems raised by the integration of processing capabilities within memory units, IMC also involves the development of new frameworks and tools that can effectively leverage memory-centric computing. Additional obstacles are raised by a knowledge gap among developers, a still-maturing support ecosystem, and the need to prove the reliability of these new technologies in diverse operational environments.

Despite these challenges, the potential benefits of IMC have spurred considerable interest and investment in the technology, with the paper “The Landscape of Compute-near-memory and Compute-in-memory: A Research and Commercial Overview”, authored by A. Khan et al., estimating rapid and sustained growth over the next decade, with many startups and established companies racing to develop and commercialize their solutions, driven by the increasing demand for faster and more energy-efficient computing. As the technology matures and more turnkey solutions become available, IMC is poised to become a cornerstone of the next generation of edge computing infrastructure.

To learn more about how Synthara’s ComputeRAM™ is unifying the embedded computing landscape by seamlessly integrating IMC into general purpose chips, allowing them to be used for a wide range of edge-AI and DSP applications, such as wearables, IoT devices, robots and smart sensors, read our upcoming blog on this topic.