What is the future of semiconductors?
The new crown epidemic has impacted the global supply chain. After some people first understood the reasons for the shortage of chips, they found that they could install more than 2 trillion transistors on a chip through continuous attempts, and they also found that the future of Moore's Law is to a large extent. Depends on where you put the wires, how small you make the transistors, etc.
When and how will the chip shortage end?
It often happens to some automakers that even though they have responded to the risk by adding multiple suppliers, those suppliers or their upstream companies may all be using products from the same small group of semiconductor fabs.
Looking back, automakers panicked when the outbreak hit, canceling orders. Then, when people still seem to want to buy a car, they find that all the display drivers, power management chips and other low-margin stuff they need have been caught up in the "work, study and live from home" consumer boom. When automakers plan to repurchase chips, they find that the long line is almost a year later, causing panic again.
Chipmakers struggled to keep up with demand and launched a rapid wave of expansion, albeit mostly in higher-margin advanced-process chips, not those matured in the automotive industry. The latest data from the chip manufacturing equipment industry association SEMI shows that sales of semiconductor equipment will exceed $100 billion in 2021, which is unprecedented.
And automakers have learned their lesson. In the summer of 2021, GlobalFoundries Fab 8 in Malta, New York, was linked to the automotive electronics industry, and GF signed agreements with Ford and BMW.
Next-gen chips will be powered by transistors below
Transistors can be made as small as you want, but if you can't connect them to each other, there's no point. So Arm and the Belgian institute Imec spent several years looking for the connection. The best solution they found was to bury the interconnects that carry power to logic circuits (rather than data) into the surface of the silicon chip and have them connect to the power distribution network on the back of the chip. This research trend is suddenly in the news because Intel seems to have said something like, "Oh yeah, we're definitely going to do this in 2025."
Cerebras' new AI chip adds 1.4 trillion transistors
What has 2.6 trillion transistors, consumes 20 kilowatts, and has enough internal bandwidth to play a billion movies on Netflix? Of course, this is the second generation of the largest chip in history. Last April, Cerebras Systems unveiled a version that uses more advanced chip-making technology, surpassing previous AI processors. As a result, the chip's memory has more than doubled to 40GB, the number of processor cores has increased from the previous 400,000 to 850,000, and there are 1.4 trillion more transistors.
As incredible as it all feels, figuring out what it does is the most important thing. Cerebras demonstrated a method that computers can use to house a second-generation wafer-scale engine, the Wafer ScaleEngine, to train neural networks with up to 120 trillion parameters. For reference, the massive GPT-3 natural language processor 175 billion parameter neural network. Furthermore, today one can connect to 192 of these computers simultaneously.
Of course, Cerebras' computers aren't the only ones used to handle the massive AI training job. SambaNova is also pursuing this, and apparently Google is also looking at some very large neural networks.
IBM launches world's first 2nm node chip
IBM claims to have developed a 2nm node chip and expects to enter production in 2024. With this in mind, leading chipmakers TSMC and Samsung are going all-in on 5nm node chips, possibly testing the waters for 3nm node chip development in 2022. As mentioned last year, what most people refer to as a technology process node has nothing to do with the size of any part of the transistor it is built from. Therefore, whether IBM's R&D progress is superior to its competitors will depend on a combination of density, power consumption and performance.
What really matters is that IBM's process is yet another nod to nanosheet transistors as the future of silicon. Major chip makers are moving from today's FinFET designs to nanosheets at their own pace, and the arrival of the nanosheet era is inevitable.
RISC-V star on the rise among global chip developers
The news isn't all about transistors, processor architecture is also increasingly important. The brain of a smartphone may be based on the Arm architecture, and the laptop and the server it is connected to may be based on the x86 architecture. But a group of fast-growing companies, especially those based in Asia, are looking to an open-source chip architecture called RISC-V. What makes it unique is that it allows startups to design custom chips without paying expensive licensing fees for proprietary architectures.
Even big companies like Nvidia are deploying RISC-V, and Intel predicts RISC-V will boost its foundry business. In an increasingly polarized technology environment, Chinese companies see RISC-V as a possible path to independence and are particularly bullish on it. Alibaba also said it will provide source code for its RISC-V core.
New silicon photonics chip 1,000 times faster than transistors
While some types of optical computing are on the way, it's likely that the kind of computer switching that researchers in Russia and IBM described last October will only happen in the distant future. Relying on substances such as exciton-exciton polaritons and Bose-Einstein condensates, the device switches at about a trillion times per second. The switching is so fast that light can only travel about a third of a millimeter before the device switches again.
New DRAM could accelerate the development of artificial intelligence
A big problem with AI is that its data travels too far. Of course, this distance is in millimeters, but it's a long distance now. Now, engineers have come up with a number of ways to shorten that distance.
Instead of making a DRAM with a silicon transistor and a metal capacitor on top, a second transistor is used as a capacitor, built on top of a silicon wafer of oxide semiconductor at the same time. Two research groups have shown that these transistors can hold data for longer than regular DRAM and can be stacked in layers on silicon, making the path between the processor and important data much shorter.
Intel announces big processor architecture changes
Last August, Intel announced the company's biggest processor architecture advancement in a decade, including two new x86 CPU core architectures -- directly named Performance Core (P-core) and Efficiency Core (E-core). These cores are integrated into Alder Lake, a "performance hybrid" family of processors that includes new technology that enables the upcoming Windows 11 operating system to run the CPU more efficiently.
“This is a great time to be a computer architect,” Intel senior vice president and general manager Raja Koduri said at the time. Data center workloads have become larger, more complex, and more diverse than ever before, and how will the architecture meet the demand for higher computing performance.”
U.S. takes strategic steps toward onshore electronics manufacturing
U.S. lawmakers, concerned about Taiwan and South Korea, the only places capable of making state-of-the-art logic chips, set out to push for cutting-edge chip manufacturing in the United States. TSMC, Samsung and Intel have already started investing heavily in chip factories. Of course, China and South Korea are also making major domestic investments, as are Europe and Japan.
It's all part of a broader global economic and techno-nationalism, says Abishur Prakash, a geopolitical futurist at the Centre for the Future of Innovation in Toronto. Some argue that these “geopolitical shifts are short-term, like a byproduct of the Covid-19 pandemic, and will calm down at some point,” he said in IEEE Spectrum in May. Wrong. Countries are going in a direction that is difficult to succeed.”
How will the camera chip develop next?
A camera chip made by French startup Prophesee and major imager maker Sony that, unlike regular imagers, doesn't capture frame after frame at every tick of the clock. Instead, it only records changes in the scene. This means, lower power consumption and faster response time.
