NVIDIA, the first name that comes to mind when it comes to graphics cards, also spends time in many areas while designing advanced silicons. The technology giant is looking for ways to improve the chip design process using the silicon it produces.
The green team expects complexity in integrated circuit design to increase incrementally in the coming years. That’s why harnessing the power of GPU computing units will soon transform from an intriguing lab experiment to a necessity for all chipmakers.
NVIDIA’s chief scientist and senior vice president of research, Bill Dally, at this year’s GPU Technology Conference (GTC) behind modern GPUs and other SoCs He talked a lot about using GPUs to speed up various stages of the design process. NVIDIA believes that some tasks will be performed better and faster using machine learning rather than having humans do it manually.
Dally leads a team of 300 researchers working hard to make GPUs faster and faster. This team wants to use GPU capabilities to overcome technological challenges, automate and accelerate a variety of tasks outside of traditional methods. The said research team was 175 people in 2019, and now it continues to grow.
Bill Dally says that when it comes to accelerating chip design, NVIDIA has identified four areas to leverage machine learning techniques. For example, mapping where power is used on a GPU is an iterative process that takes three hours in a traditional CAD tool. However, this task only takes minutes when using a specially trained artificial intelligence (AI) model. After the model is taught, this timing can go down to seconds. Of course, margins for error are important in artificial intelligence. But Dally says NVIDIA’s tools have already achieved 94 percent accuracy, which is still a respectable number.
Circuit design is a labor-intensive process that requires engineers to change the layout several times after running simulations on partial designs. Therefore, training AI models to make accurate predictions about the correct parazidollarser can help eliminate a lot of manual work involved in making the minor adjustments needed to meet desired design specifications. NVIDIA can leverage GPUs to predict series of parasites using graphics neural networks.
One of the biggest challenges in designing modern chips is routing congestion, Dally says. This defect manifests itself in a particular circuit layout where the transistors and the many small wires connecting them are not optimally placed. We can compare it to a traffic jam, but instead of cars, let’s think of a series of binollars. Using a graphical neural network, engineers can quickly identify problem areas and adjust their placement and orientation accordingly.
In these scenarios, NVIDIA tries to use artificial intelligence instead of human-made chip designs. Rather than embarking on a labor-intensive and computationally expensive process, company engineers can create a surrogate model and quickly evaluate using artificial intelligence. The tech giant also wants to leverage artificial intelligence to design the most fundamental aspects of transistor logic used in GPUs and other advanced silicon.
The GPU manufacturer has already started taking the necessary steps to move to a more advanced manufacturing technology where thousands of standard cells have to be replaced according to complex design rules. A project called NVCell is trying to automate this process as much as possible with an approach called reinforcement learning.
Trained AI model is tasked to correct design errors until completion. NVIDIA claims to have achieved a 92 percent success rate to date. In some cases, cells designed with artificial intelligence can be even smaller than those made by humans. Any breakthroughs made can improve the overall performance of the design, while helping to reduce chip size and power requirements.
Semiconductor process technologies are rapidly approaching the theoretical limits we can reach with silicon. On the other hand, as the production technique changes, the cost of dollars increases. Therefore, any small improvement in the design phase can yield better efficiency, especially if reducing the wafer size.
As you know, the green team outsources chip production to companies like Samsung and TSMC. Dally, one of NVIDIA’s leading names, says that things are progressing faster thanks to NVCell. A team of ten engineers can run their jobs faster using GPU acceleration capabilities. In this way, important minds within the company can more easily focus on other areas.
NVIDIA isn’t the only company turning to artificial intelligence when it comes to chip design. Another tech giant, Google, is using machine learning to develop accelerators for artificial intelligence tasks. Google has discovered unexpected ways to optimize the performance and power efficiency of AI. Samsung’s semiconductor division, on the other hand, uses a Synopsys tool called “DSO.ai”, which has been gradually adopted by other companies large and small.
The automotive industry has suffered greatly in the last two years due to the deficiencies in semiconductor production. In this context, foundries can leverage artificial intelligence fabrication chips in mature fabrication techniques (12 nm and larger) to address the deadlock. On the other hand, since the semiconductor space is extremely competitive, most manufacturers are reluctant to invest in it.
More than 50 percent of all chips are designed with mature semiconductor processes. International Data Corporation analystsdollarseri expects this share to increase to 68 percent by 2025. Synopsis CEO, Aart de Geus, believes he can design chips using artificial intelligence in the company’s cool cars, home aledollar series and many other devices where performance is not a priority. This approach costs less than switching to a more advanced production technology. In addition, placing more chips on each wafer (silicone disc plate) again saves costs.
As you know, there are many opinions that artificial intelligence will replace humans over time. In the chip design process we mentioned, there is no such story. NVIDIA, Google, Samsung and other companies have discovered that AI can empower humans and do the heavy lifting when it comes to increasingly complex designs. People still have to find problems to solve and decide which data helps validate chip designs. Artificial intelligence, on the other hand, can process this process much faster.