Alibaba develops chip that 'merges processor and memory'

The von Neumann bottleneck places limits on the possible speed of chip operations - but Alibaba believes it has an answer

Image:
The von Neumann bottleneck places limits on the possible speed of chip operations - but Alibaba believes it has an answer

New architecture could break the von Neumann bottleneck

Chinese tech giant Alibaba has allegedly developed the world's first, DRAM-based 3D chip that merges logic and memory.

As reported by Chinese news outlet Pandaily, Alibaba Cloud announced the chip in a (now deleted) post on Chinese social networking site Weibo post on Friday.

The company said its DAMO Academy (standing for Discovery, Adventure, Momentum and Outlook) has successfully developed a 3D stacked in-memory computing (IMC) chip that can help overcome the von Neumann bottleneck, a limitation in chip speed resulting from the CPU needing to wait for data to be delivered from memory.

Alibaba said the new chip meets the needs of AI and other scenarios for large bandwidth, high capacity memory and extreme computing power. In one specific AI scenario it tested, the chip improved regular performance by more than 10 times.

Alibaba told The Register that the breakthrough chip exists, although it declined to provide additional details.

For the past 70 years, computers have been designed according to the von Neumann architecture.

In this architecture, programmes and data are held in a memory unit, and the CPU works as a separate unit. Because the memory and processor are separated, data needs to be moved back and forth between the two, making latency unavoidable.

With the development of advanced processors that can operate much faster than data can be fed to them, this model is facing ever-greater challenges.

While the computing power of processors is increasing at a rate of about 3.1 times every two years, memory performance is only rising at a rate of about 1.4 times every two years.

Alibaba Cloud believes that IMC is the best way to solve these problems. The architecture works in a similar fashion to the human brain, storing data in RAM and processing it in parallel. This substantially decreases data handling and improves parallelism and computing efficiency.

The research results from DAMO Academy have been included in the International Solid-State Circuits Conference (ISSCC) 2022, a global forum for presenting advances in solid-state circuits and systems-on-a-chip.

In the future, Alibaba believes this new architecture could be used in virtual/augmented reality, astronomical data calculation, unmanned driving, remote sensing data analysis and other scenarios.