How far is the ultra-large chip from commercialization, and will Moore’s Law in the chip industry fail? Giants do chips, do small companies still have a way to live?
Recently, exciting milestones in the chip industry have continued. Today, researchers from MIT published a paper in Nature, claiming that they have created a 16-bit microprocessor consisting entirely of carbon nano-transistors (carbon nano-transistors are considered the preferred alternative to the original chip materials – silicon materials). . The microprocessor contains up to 14,000 transistors. The researchers expect it to go commercial in less than five years. The application of this new technology will reverse the current situation of shrinking silicon transistors and the diminishing returns of the chip industry.
A week ago, at the annual chip summit, HOTCHIPS, Silicon Valley artificial intelligence startup Cerebras Systems released the largest AI chip in history. With a die area of 42,225 square millimeters and 1.2 trillion transistors, the chip is 56.7 times larger than the largest NVIDIA GPUs and is designed to handle AI applications.
Next, Intel released the first AI chip that was built in four years. Huawei also announced that its AI processor Ascend 910 has been commercialized and launched Full Scene AI calculation framework MindSpore.
The picture shows recent chip industry memorabilia
Whether it is research institutions, chip startups, or integrated circuit suppliers including Intel, Baidu, Huawei and other technology giants, many forces are embracing the innovative research and development of chips. This will bring a deep chemical reaction to the chip – a high-input, high-risk, slow-return industry. We use three questions to analyze the triple changes they will bring.
How far is the large chip from commercialization?
We know that the larger the chip size, the faster it can process information. Attempts for ultra-large chips have always been the direction of the industry’s efforts. Previously, due to the many difficulties faced, practitioners’ attempts failed without exception.
The arrival of the AI era, the market demand for big chips is once again coming out. After deep learning, people are trainingPracticing complex neural network models requires a lot of computing power, and it is difficult for a single GPU to meet the demand.
Before the very large chip, the industry adopted the scheme for off-chip transmission, that is, connecting the chips. (If NVIDIA spends billions of dollars, developed and introduced a bus and its communication protocol – NVLink), this interconnected system is not so cost-effective. Its cost is super linear and the return is sublinear. The connection of 100 GPUs may not be able to achieve the performance of 100 GPUs, but the cost is much larger than the sum of 100 GPUs. At the same time, the off-chip transmission of high-speed interconnects has significant ceilings. Once the chip is packaged, there are pins, and the number of pins is usually several thousand. The number of pins is limited.
The emergence of large chips has enabled on-chip transmission of data, which is not only cost-effective, but also the efficiency of previous interconnections can be solved.
However, this chip is facing commercial challenges and faces several major challenges:
1. Thermal issues:
The large chip is nearly 60 times larger than the current NVIDIA GPU. Although there is no news about its power consumption, it can be expected that its power consumption will not be low. If there is no cooling system, the chip will be Will burn out.
2. Yield:
The oversized chip has 1.2 trillion transistors, and despite the use of error redundancy, it is difficult to ensure that the core circuit is free of impurities.
3. Packaging is difficult:
Chip packaging is a challenging technology, and chips with poor packaging are difficult to use in harsh environments. At present, there is no packaging solution for this chip in the industry.
4. Clock synchronization:
When the operating frequency is high enough, the delay of signal propagation is also a problem.
If these issues can be resolved, it will have a disruptive impact on the entire AI industry.
Moore’s law in the chip world will be invalid?
The chip industry has been following Moore’s Law to maintain rapid growth, that is, every 18 months, the number of transistors that can be accommodated on an integrated circuit will double. How to put billions of transistors into a single chip in the most economical way is the biggest challenge facing the current chip industry. Today, transistors are shrinking to the 7nm process and are expected to reach 1 nm by 2030. The limit of the atom is 0.1 nm. Does this mean that Moore’s Law is about to slow down?
The answer is no. At present, Moore’s Law is still valid.
But in order to pursue the most advanced technology, the transistor is engraved more and more fine, the semiconductor factory needs to make more than ten layers of photomask, lithography on the semiconductor, etc. (M