IBM has developed a new computer chip for artificial intelligence that is 15 times more powerful than previous similar designs. This chip is an analog computer, a term that refers to computers that use continuous-time signals to represent data, in contrast to digital computers that use discrete states.
Analog computers were once the most powerful, but this changed in 1947 when the first transistor was developed at Bell Labs. This invention had a profound effect on the world, leading to the exponential growth of digital computers. Today, all computer chips we use are digital.
However, this digital approach has a significant drawback: the data movement between memory and the CPU, known as data bus, is a bottleneck that dominates the runtime and energy consumption. This is particularly problematic for applications like deep learning, which are very data-intensive.
IBM's new chip, named "Wizarding you Hermes analog chip", is designed to address this bottleneck. It's a multi-core analog in-memory computing chip with 64 cores. It eliminates the separation between memory and processing, meaning data processing happens in the memory itself.
This design is similar to how humans do mental calculations in our heads, with the computing happening in the network of interconnected neurons. The new IBM chip is designed to function in the same way.
The chip uses phase change memory technology, which is very good at multi-level storage. It can store more than one bit, and this is achieved by changing the size of an amorphous region inside the phase change memory.
The chip can implement more than 4 million parameters, and the plan is to scale it further to handle billions of parameters on a single analog chip. However, integrating these devices in a crossbar at a very high density is a significant challenge.
One of the biggest challenges of this technology is the readout electronics. In the case of phase change memory technology, analog to digital converters (EDC) are used to convert the analog signal back to digital. These circuits are massive and need to be scaled without drastically increasing electronics.
In terms of software, training a network on an analog chip would require retraining the network using hardware training. This makes the network more robust to analog noise. After that, it would need to go to some kind of compiler, which would map the network onto the different arrays.
Despite the numerous advantages of analog chips, there are still significant challenges to overcome, particularly in terms of accuracy and endurance. However, the development of this chip represents a significant step forward in the field of AI.
Here are the key facts extracted from the text:
1. IBM has developed a new computer chip for artificial intelligence that is 15 times more powerful than previous similar designs.
2. The chip is an analog computer, a type of computer that was widely used before the development of digital computers.
3. The first transistor was developed at Bell Labs in 1947, which led to the development of digital computers.
4. Digital computers have grown exponentially since then and are now used in all computer chips.
5. There are two main blocks in conventional computer architecture: memory and CPU, connected by a data bus.
6. The movement of data between memory and CPU is the bottleneck that dominates runtime and energy consumption.
7. The new IBM chip is a multi-core analog in-memory computing chip with 64 cores that can perform matrix multiplications.
8. The chip uses phase change memory technology, which can store more than one bit per cell.
9. The chip is designed to function like the human brain, with computing happening in the network of interconnected neurons.
10. The chip can execute all operations involved in convolutional and LSTM layers.
11. The chip was fabricated in Albany, New York, and is the result of a global collaboration within IBM.
12. The chip can implement more than 4 million parameters, but the plan is to scale it up to handle billions of parameters.
13. The main challenge in scaling up the technology is integrating the phase change memory devices at high density.
14. The readout electronics are a major challenge, as they require analog-to-digital converters.
15. The chip achieved 92.8% accuracy in a classical image classification task, the highest accuracy achieved to date using analog chips.
16. The chip is capable of 400 gig operations per second per area, making it 15 times more powerful than previous similar designs.
17. The chip is a big step forward towards building more efficient analog chips for AI.
18. Analog chips are still not widely adopted, but they have the potential to be used for inference applications.
19. Training neural networks on analog chips is challenging and may require dedicated architectures.
20. The weights on analog chips are stationary and can only be programmed once, which limits their use for training.
21. The endurance of phase change memory devices is limited, which can make training on analog chips challenging.