AI Cambrian Explosion: 2021 Predictions

By Patrick Moorhead - January 25, 2021

For the last two years, in January, I have published my predictions about the new AI silicon I expect in the coming year. Like the weatherman, I’ve gotten some details wrong but I’ve been reasonably accurate, at least directionally. So, here we go again with a look at likely 2021 advancements and possible failures. I base none of these predictions on internal or NDA information; these are just my opinions.

Figure 1: The dominant processor for data center AI in 2021 will be the NVIDIA A100 GPU. 
NVIDIA

Predictions for 2021 AI hardware

  1. NVIDIA will retain its dominant position in data center AI. In addition to commanding nearly 100% of the training market, NVIDIA will gain more traction in cloud inference processing due to its new Multi-Instance GPU. The MIG capability on the NVIDIA A100 provides CSPs with more flexibility and can reduce hardware provisioning costs. NVIDIA’s Automotive business will begin to see growth late in the year.
  2. NVIDIA will successfully acquire Arm, leveraging the licensing business model to monetize technologies Jensen Huang does not care to productize.
  3. Even if I am wrong concerning the Arm acquisition, NVIDIA will announce an Arm-based server to enable tight coupling of CPUs and GPUs. (The server may not ship until 2022.)
  4. The lure of an in-house chip design will compel most “Super Seven” hyperscale data centers to launch proprietary AI inference processors. These chips will be tailored for specific use cases and business needs, primarily impacting Intel Xeon, and will create tremendous headwinds for startups. 
  5. The Qualcomm AI100 platform’s performance and power efficiency will secure at least one hyperscale data center win, despite the trend noted above. Should this fail to transpire, Qualcomm could shut down its data center efforts.
  6. Google will launch the TPU4, which was teased in July when the company published mlPerf benchmarks that more than doubled the previous design’s results. Google will also announce a second generation of the Edge TPU; the edge is too significant for Google to miss. 
  7. Intel will land at least one major design win for Habana Gaudi, leveraging the AWS win announced in December. If this occurs, Gaudi will arguably earn the pole position to compete with NVIDIA.
  8. Intel will either release an updated Habana Goya inference chip or will quietly let Goya die, focusing instead on Xeon processors. I’d bet on the latter.
  9. Graphcore will announce at least one significant design win, perhaps Microsoft. Others that look promising include TenstorrentBlaizeSambaNova and Groq. However, 2021 is a make-or-break year for many startups.
  10. The last one is an easy one: someone will buy someone else. Seriously, startups with promising technology, such as Cerebras, SambaNova, Tenstorrent, Blaize and Graphcore, may look expensive but are quite valuable. Companies like AMD, NVIDIA, Facebook and Google can afford to pay up for a platform that may represent a durable breakthrough. 

Conclusions

In 2020, the AI Cambrian Explosion moved from the drawing boards and into data centers and edge devices. I expect 2021 to usher in scores of new chips to accelerate AI, from the startups and the large semiconductor vendors alike. The landscape for edge AI will become populated with dozens of companies with tailored platforms to handle specific models and at various performance levels, power envelopes and costs. In contrast, the data center will remain the domain of the largest semis, with NVIDIA in the lead and a few others nipping at Jensen’s heels. I will review this list of predictions next December to see how I did. 

Buckle up—2021 looks to be one heck of a ride!

+ posts
Patrick founded the firm based on his real-world world technology experiences with the understanding of what he wasn’t getting from analysts and consultants. Ten years later, Patrick is ranked #1 among technology industry analysts in terms of “power” (ARInsights)  in “press citations” (Apollo Research). Moorhead is a contributor at Forbes and frequently appears on CNBC. He is a broad-based analyst covering a wide variety of topics including the cloud, enterprise SaaS, collaboration, client computing, and semiconductors. He has 30 years of experience including 15 years of executive experience at high tech companies (NCR, AT&T, Compaq, now HP, and AMD) leading strategy, product management, product marketing, and corporate marketing, including three industry board appointments.