China has just embarked on a unique industrial effort by launching mass production of a non-binary computer chip. An initiative that could help diversify our computing ecosystem, and perhaps even lead to the emergence of new technologies.
Today, the vast majority of chips in the devices we use every day are built around so-called binary logic, implemented using transistors that can alternate between two distinct states—typically 0 and 1.
Binary: A Matter of Compromise
If binary is so prevalent today, it is primarily because it represented an ideal compromise at the dawn of computing. Conceptually speaking, it is an extremely simple approach, which greatly facilitated its implementation in early computer systems. Furthermore, this simplicity makes binary a perfect match for Boolean algebra, a mathematical system that relies entirely on true-or-false logic. This created an intuitive and elegant link between abstract mathematical logic and physical implementation, thus paving the way for the first modern computers. But even though binary logic is an extremely reliable, robust, and easy-to-handle foundation, that doesn't mean it's perfect. Today, it remains the standard because an entire industry has grown up around this approach, to the point where it's become almost impossible to part with it for both logistical and economic reasons. Yet, alternatives exist, and there are even many cases where other systems would be more efficient.
Machine learning is a good example. As a reminder, current AI models are built around artificial neural networks that can be stimulated by addressing them with a value of 1 or 0. With a ternary system capable of using three values (-1, 0 and 1 instead of 0 and 1), we could also inhibit certain neurons with a value of -1 to improve the efficiency of the inference process.
And this is only the tip of a huge iceberg; Some researchers are even convinced that it would be interesting to develop alternative approaches, for example based on a non-binary system, to advance our technological ecosystem. And this is precisely what the troops at China's Beihang University in Beijing are currently working on.
A hybrid chip based on probabilities
According to the SCMP, Professor Li Hongge's team has developed a hybrid system called HSN, for Hybrid Stochastic Number. The central idea is to use this good old binary in addition to another approach called stochastic to design a chip that is both highly efficient and exceptionally versatile.
In the binary world, each value corresponds to a very specific sequence of bits; 1000 in binary corresponds, for example, to the number 8. But you only need to change the position of one of these bits to change the nature of the information. For example, shift the first bit to get 0100, and you get 4 instead of 8.
This is both an advantage and a disadvantage. On the one hand, it is a completely deterministic approach; the same operation always produces the same result. But on the other hand, it also means that binary is both very sensitive to errors and not necessarily very efficient, because sometimes several mathematical operations must be chained together to make a simple modification.
To avoid these shortcomings, we can also use a stochastic approach, that is, one based on the notions of randomness and probability. For example, we can arbitrarily decide that the number of 1s in a sequence of 10 bits represents a digit from 0 to 9. In this case, our 8 would be represented by a series like 0110111111… but also 1111101011. Here, it is not the exact value of each bit that matters, but rather their statistical distribution.
Again, this is both an advantage and a disadvantage. The main concern is that this relatively old approach, conceptualized as early as the 1960s, is not entirely deterministic. Each individual piece of information cannot be trusted, and the true value can only emerge from a sufficiently large set of bits. This can obviously be problematic in applications that rely on pure arithmetic precision. But on the other hand, the non-dependence on absolute values makes this approach more error-resistant. It also allows certain operations to be performed much more quickly, and even in parallel.
By making the two technologies coexist and interact on the same support, the Chinese team was able to design a new type of chip. Thanks to the complementarity of binary and the stochastic method, it is able to exploit the respective advantages of the two approaches while limiting their disadvantages.
A component already in mass production
This system has already been implemented on a chip that is now in mass production by SMIC, the leading Chinese foundry. It is now used in touchscreens, a technology particularly suited to this hybrid approach.
For example, certain functions, such as calculating pixel coordinates, require exact mathematics; they are therefore handled by the binary part. But at the same time, touchscreens also produce a large amount of electronic "background noise" that absolutely must be filtered out, for example to determine whether the user is pressing on the screen or has simply touched it. This is a relatively demanding task for a binary system, but one that a stochastic system can perform very easily. With this hybrid system, both parts can be handled very efficiently by a system with low computing power.
"The current chip already achieves microsecond-level on-chip computing latency, striking a balance between high-performance hardware acceleration and flexible software programmability," the authors explain.
Is the industry on the cusp of a major shift?
And this may be just the beginning. The team is developing a set of instructions (basic commands that tell a processor what to do) specifically tailored to this hybrid stochastic approach. Once mature, this could make it possible to design extremely high-performance control systems for very specific applications, particularly in areas such as machine learning.
Admittedly, there is little chance that this approach will become widespread in the near future; Binary is too deeply rooted in our society to consider a major replacement. But it will be very interesting to see if this initiative will be emulated.
We live in an era where the physical limits that condition our computing capacity (transistor size, temperature and energy management, etc.) are becoming increasingly difficult to push back to meet the needs of our society. The closer we get to these limits, the more interesting it will become to move away from binary and favor hybrid approaches of this kind; it is not impossible that more and more companies and institutions will begin to explore this horizon, which could considerably transform our technological ecosystem. See you in a few years, or rather decades, for a new assessment!
0 Comments