8-bit analog memory for working with neural networks

    At the end of the year, an article appeared on Habré that IBM managed to create the first 8-bit chip based on phase-change memory (PCM) memory.

    Today we will tell how the device represented by the IT giant works.

    / photo IBM Research CC BY-ND

    What is the purpose of chip development?

    Modern machine learning systems are mainly based on graphics processors (GPU). They require active cooling, since GPUs can heat up to half the speed of central processing units (CPUs). This translates into large electricity bills for data centers in which such computing clusters are located.

    The issue of reducing power consumption is important for data centers. According to analystsBy 2025, data centers will consume 20% of the total generated electricity. In order to reduce data center costs, new technologies for machine learning (MO) are being developed. These include 8-bit analog phase-change memory (PCM memory) from IBM. According to the developers, the new chips will be a hundred times more productive than classic GPUs and will require 280 times less power.

    The advantages of PCM memory are due to the fact that data processing and storage occur on the same device. In traditional systems, these processes are separated - data has to be transferred from device to device, and because of this, the speed of calculations and their energy efficiency decreases.

    The chip works with 8-bit floating point numbers. Usually in machine learning they use numbers with greater accuracy, however, according to researchers from IBM, in “everyday” tasks this is enough.

    The creators of the technology managed to apply it in the translation of texts and recognition of speech and images. IBM vice president of research, Jeff Welser (Jeff Welser) compared the work of a neural network in a chip with a person who sees a vague figure in a sweaty window, but still recognizes his relative or friend in this spot.

    IBM has already shown the first results of PCM memory with machine learning technologies. A simple single-layer neural network was launched on the chip, which recognized handwritten numbers 1, 0 and 4. The algorithm distinguished images with 100% accuracy.

    How does IBM analog memory work

    The chip is made of an alloy of germanium, tellurium and antimony, placed between two electrodes. When a current is passed through the material, it changes the state from crystalline to amorphous. These states mean zero or one.

    Due to the fact that the phase does not change in the entire volume of the alloy, but only in its part, intermediate values ​​between 0 and 1 can be stored in the cell. When writing to the memory, the numbers are not “rounded”, and this helps to achieve greater accuracy of the algorithms that work on the chip. It is believed that this method of data storage more reliably conveys the behavior of neurons in the human brain.

    When reading data, a small current is passed through the cell and the resistance is measured. In the amorphous state, the resistance of the alloy is higher than in the crystalline state - by this indicator it is determined what number is stored in the cell.

    Although the reading voltage is lower than the required for changing the phase state of the material, it leads to a small shift of the particles and distortion of the data. Therefore, IBM improved the PCM-memory and added an additional layer of titanium nitride to the device, which “ wraps ” the alloy - it was called the reflector.

    / Koelmans, WW et al. Projected phase-change memory devices / SS BY

    When writing data, the reflector does not affect the passage of current through the alloy, since the resistance of titanium nitride in this case is higher than the resistance of the cell material. When reading data, the resistance of the material of the cell becomes higher than that of the reflector, and therefore the current passes through the outer layer, bypassing the alloy. This type of memory at IBM was called Proj-PCM, or projected PCM - "PCM with a reflector."

    Technology perspectives

    In the IT industry, a new chip was treated positively. According to the scientific editor of the publication Ars Technica, the device has great potential - with the development of technology and reduction in the size of the memory cell, IBM can increase the device performance at least a hundred times. However, the IT giant has yet to solve several difficulties.

    The first is the mass production of PCM memory. According to processor analyst Linley Gwennap (Linley Gwennap), the characteristics of the chips, for example, the operating voltage and the resistance of electronic circuits, vary from device to device. Gwennap believes that IBM has yet to develop a suitable manufacturing process, so the analog chip will not soon become a commercial product.

    Another difficulty is sustainability.to temperature changes. Now, under the influence of heat or cold, the alloy changes its phase. Therefore, the creators of the device will come up with a way to protect the chip from temperature fluctuations. Probably, the output will be to reduce the size of the device - studies show that the material is less susceptible to the influence of temperature with a layer thickness of several nanometers.

    / photo IBM Research CC BY-ND

    Who else is developing chips for MO

    A similar technology is developing Mythic startup. The company's device is based on a multi - level flash memory (multi-level cell, or MLC). A chip cell can store not two or four charge levels, as in more common flash memory devices, but 256. This significantly increases its capacity. It is expected that this memory will be used to implement machine learning algorithms in the Internet of things devices: smartphones, drones, smart speakers and hearing aids .

    Another similar chip createsSyntiant. Calculations in the device are also performed in the MLC-memory. The main advantage of the technology is low power consumption. Representatives of the company say that their device consumes 50 times less electricity than digital processors. Syntiant will also be used in battery-powered gadgets: smartphones, wearable devices, sensors, and surveillance cameras. The startup is already working on the first commercial implementation of the technology - a “smart” microphone.

    In general, analog PCM chips are not yet designed for high-precision computing. Therefore, they are used in combination with central and graphics processors. IBM believes that hybrid architecture can create a balance between performance and precision. But in the future it is expected that analog memorywill be able to replace classic chips in data centers.

    The first corporate IaaS blog:

    Our IaaS blog in Telegram:

    Also popular now: