IBM has developed a new computer chip for artificial intelligence, the Hermes Analog Chip, which is a multi-core analog in-memory computing chip. This chip is 15 times more powerful than previous designs and aims to address a significant issue in conventional computer architectures: the bottleneck that occurs when data is moved between memory and the CPU. This data movement, often referred to as "fetching," consumes a significant amount of energy and contributes to the chip's runtime.
The Hermes Analog Chip eliminates this bottleneck by performing computations in the memory itself, reducing the need to constantly move data between memory and the CPU. This approach is similar to how humans perform mental calculations, where computations occur within the network of interconnected neurons.
The chip uses phase change memory technology, which can store more than one bit, enabling the storage of multiple bits or the equivalent of multiple bits in a single memory cell. This is achieved by modulating the size of the amorphous region inside the phase change in every device.
The Hermes Analog Chip was fabricated in Albany, New York, and can implement more than 4 million parameters. However, the goal is to scale it further up to handle billions of parameters on a single analog chip.
The chip's performance in a classical image classification task was 92.8%, the highest accuracy achieved to date using analog chips with similar technology. The chip's throughput, measured in gig operations per second per area, is 400, which is 15 times more powerful than previous similar designs.
Despite the potential benefits of analog chips, there are challenges that need to be addressed. These include the difficulty of integrating phase change memory devices at high density, the limitations of the metal layer in the back end integrations, the readout electronics required for phase change memory technology, and the accuracy and endurance of these devices. Currently, digital chips seem to be a better fit for deep neural network training.
1. IBM has built a new computer chip for artificial intelligence.
2. The chip is 15 times more powerful than previous similar designs.
3. It's an analog computer, a type that was the most powerful until 1947.
4. The invention of the transistor at Bell Labs had a profound effect on the world, leading to the exponential growth of digital computers.
5. Conventional computer architectures often have two main blocks: memory and a CPU, connected by a data bus.
6. The data movement between memory and CPU is a bottleneck that dominates runtime and energy consumption.
7. The IBM New Hermes Analog chip is designed to address this bottleneck.
8. It's a multi-core analog in-memory computing chip with 64 cores.
9. The chip eliminates the separation between memory and processing, performing computations in memory.
10. It aims to reduce energy consumption by moving the compute engine to the memory.
11. IBM's chip is based on phase change memory technology, which can store more than one bit.
12. The chip can implement more than 4 million parameters and is designed to scale up to handle billions of parameters in the future.
13. The chip was fabricated in Albany, Upstate New York, and the phase change memory is done by an external foreign company.
14. The chip can achieve 400 gig operations per second per area, making it 15 times more powerful than previous similar designs.
15. While there is a lot of potential in analog chips, it's still unclear if they will be widely adopted for training, which requires flexibility and customization not easily achievable with analog architectures.
16. Analog chips are more suited to inference applications due to the limitations in reprogramming weights during training.
17. The endurance of phase change memory devices is a limiting factor for training, as they can only be reprogrammed a certain number of times before they die.