No matter which market segment a startup decides to join, competition will be fierce. I find it useful to map this market as a triangle, as shown in Figure 1, with each corner having its own set of criteria to represent unique market needs.
The top corner of the triangle is the demand for AI chips in data centers, cloud and high-performance computing (HPC) environments. Cerebras is a new company that has mastered this market and created what is claimed to be the world’s largest chip – Wafer Scale Engine (WSE). This market segment requires the highest computing power, followed by power consumption and cost. The challenge for new startups here is to compete with hyperscalers or established players – Nvidia, for example, continues to steadily improve its architecture, and the latest version (Ampere) was released in May 2020.
The base corners of the triangle are primarily relevant for corollary, allowing for the creation of wafers that maintain accuracy but reduce precision, subject to varying constraints including wafer size, low latency, low power consumption, and low unit cost. The small edge device market is the most active area for new startups, and big-name players like Nvidia are less likely to compete. The company has also stated that it does not plan to enter the mass commodity inference market. But players in this space not only have to compete with rivals, but also with potential customers who may decide to start or acquire a new company.
Figure 1: Segmented application markets of AI hardware accelerators.
(Image source: Kiasco Research)
The next development of the AI chip market
There are too many competing companies in the field of AI chips. If you look at each corner of the triangle in Figure 1, you will know that only the best all-round design will win. In addition to the various factors we have already mentioned, a mature software development stack, a vision for the market, and a greater potential to embed deep learning applications into products need to be added to rationalize the market.
Some people have already been “dead” in this market: the most recent case is that Wave Computing declared bankruptcy in April 2020.
Competition is driving the market for faster, higher-performance AI chips, from which AI researchers will benefit and realize their innovative designs. The author also expects that new algorithms will emerge to replace the current dominance of deep learning – the long-term vision of AI research is to create an artificial brain similar to the human brain. Obviously deep learning will reach a dead end. The emergence of new algorithms is inevitable (some have already appeared, but are beyond the scope of this article), and these new generations of algorithms may require different types of accelerators.
The breadth of deep learning applications in real-world cases has given these chips a multibillion-dollar market that will continue to grow with the rollout of 5G. This market needs AI hardware accelerators, the AI chip market will tend to be rationalized, and then the entire rules of the game will change with the rise of a new generation of AI algorithms… However, no one can say when the change will happen.
