Collaboration looks to build AI Accelerator chip with analogue in-memory computing

Find the Latest SoC Solutions
for... Automotive... IoT... Security... Audio... Video

Andes Technology, a supplier of 32/64-bit RISC-V processor cores and TetraMem, a developer of analogue memristor technology and in-memory computing, are collaborating on the development of a fast, highly efficient, AI inference chip.

www.newelectronics.co.uk/, Aug. 15, 2023 – 

TetraMem has licensed the Andes RISC-V NX27V vector CPU, combined with ACE (Andes Custom Extension) to create a solution that addresses the challenges of AI processing in power-constrained environments.

At the heart of this collaboration is the fusion of Andes' high-performance RISC-V Vector CPU with TetraMem's compute memristor - an analogue RRAM - in-memory computing architecture through ACE to enable tight coupling and to improve the device's performance. This combination amplifies the strengths of both companies, resulting in a much faster, energy-efficient AI inference that surpasses the limitations of traditional computing approaches.

Andes RISC-V Vector CPU cores provide performance, efficiency, and configurability, making them suitable for a wide range of AI and edge computing applications and the addition of Andes powerful vector processor brings unparalleled performance capabilities to the accelerator chip.

TetraMem's analogue in-memory computing technology empowers the chip with massively parallel VMM computation without data movement, mitigating the energy overhead of conventional architectures as confirmed in TetraMem's first commercially manufactured demonstration chip.

The joint effort aims to create a chip that is not only powerful but improves energy-efficient by at least an order of magnitude. By optimising computations and eliminating transfer of weight data, the planned chip will significantly extend the battery life of edge devices and impose a near-zero impact on thermal budgets.

The AI accelerator chip will be designed from 22nm and beyond, to 7nm and below in the future, with a focus on versatility and scalability ensuring easy integration into various AI-powered products and applications. This adaptability ensures broad industry applicability. The TetraMem founding team has demonstrated scalability of the compute memristor to 2nm and below, ensuring a roadmap to future-proof solutions.

click here to read more...

 Back