With 80 startups and 34 established players, the AI chip market is clearly unsustainable in its current form. Why is it this way, how will it change, and what does this all mean?
Right now, the AI chip market is all about deep learning. Deep learning (DL) is the most successful of machine learning paradigms at making AI applications useful in the real world.
The AI chip market today is all about accelerating deep learning (DL) – the acceleration is needed during training and during inferencing. The AI chip market has exploded with players: for a recent research report we counted some 80 startups globally with $10.5 billion spend by investors, competing with some 34 established players. Clearly this is unsustainable, but we need to dissect this market to better understand why it is the way it is now, how it is likely to change, and what it all means.
Winding the clock back to around 2010 when Nvidia launched its high-end general-purpose computing on graphics processing units (GPGPUs) – we just call them GPUs now – it led to the rise of DL, reducing training times of large neural networks from months and weeks to days and hours, and less. Nvidia grew a new multi-billion-dollar business around being the AI computing company. That prompted other chip companies and chip architects to think about how they could build an architecture dedicated to running AI workloads, starting with a clean sheet, and do better than GPUs designed for more varied workloads. The AI workloads today simply mean running DL, this is where the market demand exists.
But the market is varied in its needs. While most of the AI training is performed in the data center (including on hyperscale clouds) and on workstations, the AI inferencing is done everywhere: on the cloud, on the workstation, on the edge. Especially the edge.
How the AI chip market segments itself
Whichever segment of the market a startup decides to address, the competition will be tough. I found it useful to map the market as a triangle, see Figure 1, and each vertex represents a distinct market demand with its own set of criteria. At the apex is the AI chip demand in data centers, clouds, and HPC environments. Cerebras nicely captures addressing this market: it has built the largest chip on the planet, its wafer scale engine. This market segment wants maximum compute performance, and power usage is a secondary factor, as is cost. The challenge for the startups is that they are competing with the hyperscalers and incumbents – Nvidia has a steady release cadence of improved architectures, the latest (Ampere) was launched in May 2020.
At the lower edge of the triangle, where it is mostly about inferencing, it is possible to build chips with lower precision while retaining accuracy. The constraints are now different: size of chip, low latency, low power consumption, and low cost per unit. The small edge in particular is the most active across startups. The competition here is less from incumbents like Nvidia, which has said it is not planning to play in the mass commodity inference market. But the players in this corner not only battle their rivals but also their potential customers, who may decide to build their own or buy up a startup.
What comes next for AI chips
We said there are too many competing players in the AI chip space, and we are looking at each Figure 1 vertex separately. The best all round designs will win out: to the various factors we have already mentioned there needs to be added a mature software development stack, and a vision for addressing the market and its wider potential for embedding DL applications in products. Market rationalization will take place. There have already been casualties: the latest is Wave Computing declaring bankruptcy in April 2020.
The market competition is leading to faster, higher performance AI chips on the market. AI researchers will reap this benefit to run their novel designs and I expect new algorithms will emerge that will overtake DL’s current supremacy. For long term AI research with aspiration to build human-like artificial brains, it is clear that DL is a dead end. Inevitably, new algorithms will emerge (some already have, but that is another article), and these next-generation algorithms may need a different type of acceleration.
The breadth of DL applications in real-world use cases is such that there is a multi-billion-dollar market for these chips, set to grow with 5G rollout. This market needs AI hardware accelerators. The AI chip market will rationalize and then the whole game will change over as next-generation AI algorithms take the lead, but no one can say what the timing for that will be.
– Michael Azoff is Chief Analyst at Kiasco Research