IBM Develops New 64-Core Hybrid Analog Memory Computing Chip

2023-08-28

For decades, electrical engineers have been striving to develop advanced devices that can perform complex calculations faster and consume less energy. With the emergence of artificial intelligence and deep learning algorithms, this need has become even more urgent, as these algorithms often have high requirements in terms of data storage and computational load.

One promising approach to running these algorithms is called Analog In-Memory Computing (AIMC). As the name suggests, this approach involves developing electronic devices that can perform calculations and store data on a single chip. To achieve improvements in speed and energy consumption, this approach ideally should also support digital operations and communication on the chip.

Researchers at IBM Research Europe have recently developed a new 64-core mixed-signal in-memory computing chip based on phase-change memory (PCM) devices, which better supports the computation of deep neural networks. They described this 64-core chip in a paper published in the journal Nature Electronics, and achieved very promising results, preserving the accuracy of deep learning algorithms while reducing computation time and energy consumption.

One of the authors of the paper, Manuel Le Gallo, told Tech Xplore magazine, "We have been studying how to perform computations with phase-change memory (PCM) devices for over 7 years, starting from our first demonstration of how to implement neuron-like functionality with a single PCM device." "Since then, we have demonstrated many applications that can benefit from using PCM devices as computational elements, such as scientific computing and deep neural network inference. With this new chip, we hope to take a step towards an end-to-end analog AI inference accelerator chip."

To create their new in-memory computing chip, Le Gallo and his colleagues combined PCM-based cores with digital computational processors, connecting all the cores and digital processing units through a digital communication network on the chip. Their chip consists of 64 analog PCM cores, each containing a 256×256 crossbar array.

Le Gallo explained, "We integrated compact time-based analog-to-digital converters in each core for conversion between the analog and digital worlds." "Each core is also integrated with lightweight digital processing units for executing rectified linear unit (ReLU) neuron activation functions and scaling operations. In the middle of the chip, there is also a global digital processing unit for executing long short-term memory (LSTM) network operations."

One unique feature of the team's chip is the connection between the internal memory cores and the global processing unit through a digital communication network. This allows the chip to perform all the computations related to the layers of a neural network on the chip, greatly reducing computation time and power consumption.

To evaluate their chip, Le Gallo and his colleagues conducted a comprehensive study, running deep learning algorithms on their chip and testing its performance. The evaluation results were very promising, achieving an impressive accuracy of 92.81% in training a deep neural network for image recognition when running and testing on the CIFAR-10 image dataset.

Le Gallo said, "We believe this is the highest accuracy among all chips using similar techniques." "In the paper, we also demonstrated how to seamlessly integrate analog in-memory computing with multiple digital processing units and digital communication structures. The chip has an area throughput of 400 GOPS/mm2 for 8-bit input-output matrix multiplication, which is more than 15 times higher than previous multi-core in-memory computing chips based on resistive memory, while achieving comparable energy efficiency."

The latest work from IBM Research Europe is a further step towards the development of AIMC chips that support the requirements of deep learning algorithms. In the future, the chips designed by Le Gallo and his colleagues may undergo further updates to achieve even better performance.

Le Gallo added, "Based on our experience with this chip and another 34-chip design published earlier this year in VLSI, we have designed an end-to-end analog AI inference accelerator architecture, which was published in IEEE's VLSI Systems earlier this year." "Our vision is to mix many analog in-memory computing chips with a specialized digital computing core, connected through a large-scale parallel 2D grid. Combined with our recent developments in hardware-aware training, we expect these accelerators to provide software-equivalent neural network accuracy on various models in the coming years."

RECENT AI NEWS

RECENT AI TOOLS