Samsung: World’s first MRAM to pave the way for next-generation AI chips

Samsung: World’s first MRAM to pave the way for next-generation AI chips

Samsung in-memory computing on Nature: the world’s first MRAM to pave the way for next-generation AI chips

According to a report on January 17, Samsung Electronics recently published the world’s first research on in-memory computing based on MRAM (Magnetic Random Access Memory) in the top academic journal Nature.

In-memory computing greatly reduces the power consumption of AI computing because it does not require data to be moved between the memory and the processor, and is regarded as a cutting-edge research in edge AI computing.

Although MRAM memory devices have the advantages of durability and mass production, their low resistance properties prevent this type of memory from being used for in-memory computing.

This time, the Samsung Electronics research team built a new MRAM array structure and used the MRAM array chip based on the 28nm CMOS process to run AI algorithms such as handwritten digit recognition and face detection, with an accuracy rate of 98% and 93%, respectively.

Investigate extended in-memory computing types to fill MRAM gaps

In traditional computer architecture, data needs to be moved from memory to processing units, and intermediate results are then stored back in memory.

But this unnecessary information transmission not only increases the computational delay, but also increases the associated power consumption.

According to TSMC’s previous research on in-memory computing, the energy consumption of data movement is even greater than that of computing.

Therefore, in-memory computing, which performs data storage and computation simultaneously in memory, has become the focus of research in both industry and academic institutions.

In previous studies, non-volatile RRAM (resistive random access memory) and PRAM (phase change random access memory) are the two most commonly used types of memory for in-memory computing.

In contrast, despite the advantages of MRAM devices in terms of operating speed, endurance, and mass production, their lower resistance makes MRAM memory unattainable in traditional in-memory computing architectures for low power consumption.

In this paper, Samsung Electronics researchers built a new MRAM-based in-memory computing architecture to fill this gap. The paper specifically writes that the research does not compete with other memory-based in-memory computing architectures.

So far, no single memory type has dominated electronics, as each type of memory has its own advantages and disadvantages. Therefore, in-memory computing based on different memories may also develop into different architectures.

From this perspective, Samsung Electronics contributes to the development of in-memory computing by filling the gap in the in-memory computing architecture based on MRAM memory.

▲ Samsung research team: Co-corresponding author Donhee Ham (first from left), academician of SAIT (Samsung Advanced Institute of Technology) and professor at Harvard University; Seungchul Jung (second from left), researcher of SAIT who wrote the first paper; third from left)

Accurately detect 1851 faces based on 28nm CMOS process

Samsung Electronics built a 64×64 MRAM array, and its peripheral circuits were integrated through a 28nm CMOS process.

Specifically, the MRAM array is between the write/read (W/R) electronics and the TDC readout electronics at the bottom, with the input data controller (IN) on the left side of the array.

In order to make up for the problem of MRAM’s small resistance, Samsung Electronics introduced a new bit cell (element at the intersection of the array row and column, bit-cell), each bit cell is combined into two paths, each path is composed of an MTJ (magnetic tunnel junction) and a FET (field effect transistor) switch in series.

▲ Chip layout and MRAM array layout (Image source: Nature)

The researchers connected these new bit cells in series in each column, and the column resistance R was obtained by adding the output resistance of each bit cell.

Through the new structure design, the column resistance R replaces the column current sum based on Kirchhoff’s law in the traditional (in-memory computing) structure and becomes the column output, which solves the problem of small resistance of MRAM devices.

Simply put, Samsung Electronics has developed an MRAM array chip that replaces the standard “current sum” memory computing architecture with a new “resistance sum” memory computing architecture, thus solving the small resistance problem of a single MRAM device.

To use this new MRAM array for AI computing, the researchers employed a binary neural network (BNN) algorithm.

The accuracy of such an algorithm can either represent each real-valued weight as a binarized function at the cost of network size, or represent each real-valued input data as a sequence of multiple binarized functions at the cost of computational speed, to Improve the accuracy of the BNN algorithm.

In this study, Samsung Electronics adopted the latter approach, expanding each input data into 8-bit codes to suppress noise. After that, the researchers used a two-layer BNN network to classify the MNIST dataset.

The MNIST dataset is from the National Institute of Standards and Technology (NIST), and the training and test sets are composed of different handwritten digits. 50% of the handwritten digits in the dataset and test set are from high school students, and the other 50% are from the U.S. Census Bureau. staff of.

▲ Different handwritten numbers 7 (Image source: CSDN)

The researchers performed the classification of 10,000 images of handwritten digits with the MRAM array, repeated three times, and achieved an accuracy of 93.23±0.05%.

After the test, the researchers classified 10,000 handwritten digit images through the VGG-8 neural network, and the accuracy rate was as high as 98.86±0.06%.

In addition to handwritten digit classification, Samsung Electronics runs face detection algorithms with four MRAM array chips.

In this step, the MRAM array chip does not perform face authentication on all objects in the scene, but first detects whether there is a face in the scene, and then activates the higher-power face authentication algorithm after confirming the existence.

Through this method, the MRAM array chip detected 1851 faces from 2000 unmasked face scenes with an accuracy of 92.5%; 483 faces were detected from 500 masked face scenes with an accuracy of 92.5%. 96.6%, with an overall accuracy of 93.4%. In addition, MRAM array chips can be combined with cameras to detect faces in real time.

▲ 4 MRAM array chips are connected to the camera for real-time face detection (Image source: Nature)

Integrating multiple devices faces challenges and may be used in biological neuron networks in the future

For the study, the researchers wrote that an important challenge for MRAM arrays to run in-memory computing is building AI SoCs (systems on a chip) that integrate many arrays with data converters, digital electronics.

The researchers also emphasized that, broadly speaking, memory arrays can be used not only to operate neural network algorithms, but also as potential carriers of biological neuron networks.

In September 2021, Samsung Electronics and Harvard jointly published a paper titled “Neuromorphic electronics based on copying and pasting the brain” in Nature Electronics, a sub-journal of Nature, proposing a The possibility of “copying and pasting” the brain’s neuronal wiring map onto a high-density 3-dimensional storage network.

▲ Samsung’s previous “copy and paste” brain research (Image source: Nature)

According to Seungchul Jung, first author of the MRAM array study, in-memory computing is similar to computing in the human brain because human computing also occurs in memory or synaptic networks.

While the current computing purpose of MRAM arrays is not to mimic the brain, this solid-state storage network may in the future be used as a platform for simulating brain synapses.

Conclusion: New research may enrich Samsung’s in-memory computing products

In recent years, in-memory computing has gradually become a common knowledge in the industry and academia. Relevant papers have repeatedly appeared in top conferences in the electronics field such as ISSCC and IEDM, and leading semiconductor manufacturers such as TSMC are also planning and exploring.

As a global storage leader, Samsung Electronics has been paying attention to in-memory computing technology.

In February last year, Samsung Electronics announced its first high-bandwidth memory (HBM) integrated with AI computing power, which can save more than 70% of energy consumption and provide more than 2 times the system performance.

This MRAM-based in-memory computing research has enriched Samsung’s layout in the field of in-memory computing, and we may be able to see more types of in-memory computing products appearing on the market in the future.

Samsung: World’s first MRAM to pave the way for next-generation AI chips