On Friday, Alibaba Cloud announced in a social media post that its DAMO Academy has successfully developed a 3D stacked In-Memory Computing (IMC) chip.
Alibaba Cloud claims this is a breakthrough that can help overcome the von Neumann bottleneck, a limitation on throughput caused by the standard personal computer architecture. It meets the needs of artificial intelligence (AI) and other scenarios for high bandwidth, high capacity memory and extreme computing power. In the specific AI scenario tested by Alibaba, the performance of the chip is improved by more than 10 times.
With the outbreak of AI applications, the shortcomings of the existing computer system architecture are gradually revealed. The main problems are that, on the one hand, processing data brings huge energy consumption. Under the traditional architecture, the power consumption required for data transmission from memory unit to computing unit is about 200 times of that of computing itself, so the real energy consumption and time used for computing are very low.
On the other hand, the development of memory technology lags far behind that of the processor. At present, the computing power of the processor is increasing at a rate of 3.1 times every two years, while the performance of the memory is only improved by 1.4 times every two years. The performance of data storage technology greatly affects the speed of data transmission.
Alibaba Cloud believes that IMC is currently the best way to solve the above problems. Similar to the human brain, IMC stores data in RAM and processes it in parallel, supplying real-time insights that enable businesses to deliver immediate actions and responses.
This technology was proposed as early as the 1990s. However, due to the complexity of technology, high design cost and lack of application scenarios, the research progress of IMC in the industry has been slow in the past decades. Nowadays, Alibaba‘s DAMO Academy hopes to solve the industry problem of computing power bottleneck through self-developed innovative technologies.
In addition, the IMC has broad application prospects. For example, in VR/AR, unmanned driving, astronomical data calculation, and remote sensing image data analysis, IMC can give full play to its features of high bandwidth and low power consumption.
At present, the research results of this chip have been included in the International Solid-State Circuits Conference 2022, the foremost global forum for presentation of advances in solid-state circuits and systems-on-a-chip. In the future, DAMO Academy hopes to make further breakthroughs in IMC technology and gradually realize typical applications to help build the technological ecosystem.