The biggest news in semiconductor over the past week is not from any semiconductor company but our usual suspects at Cupertino and Seattle.
Apple's growing ambition in silicon
Shortly after getting very positive feedback for the M1 chip, Apple is reported to be working on several successors to power the higher end Macbook or even the Mac Pro workstation as early as spring of 2021. Apple initially expected the transition to Apple Silicon to complete by 2022. If the latest source is credible, it seems like the transition is well ahead of schedule.
In a separate development, Apple's senior VP of hardware technologies, Johny Srouji, disclosed in a town hall meeting that the company had started developing its own cellular modem this year to replace Qualcomm. For everyone following Apple's silicon effort, this is likely not a surprise. Apple spent $1 billion acquiring Intel's smartphone modem business last year, and developing its cellular modem is just a matter of time. Honestly, we are a bit surprised that Apple didn't start earlier. However, designing a cellular modem is a very different and challenging task than developing an ARM CPU. We note that Apple and Qualcomm signed a multi-year supply agreement for smartphone modem in early 2019, and we don't expect any change to that in the near term.
Adopting machine learning models will become more affordable in the coming years
In the annual AWS developer conference, the company introduced another custom machine learning processor called Trainium, the second in-house machine learning-focused processor from AWS after launching Inferentia two years ago. As the name Trainium suggests, the new chip extends into the ML training workload to lower the cost of ramping AI models for AWS customers.
Alphabet is the first major cloud vendor to start developing custom processor for machine learning with the introduction of TPU, and Amazon is quickly catching up. Both companies can now address both machine learning training and inference workloads with the internally developed processors. We expect other major cloud vendors to follow the same path. With the development of internally designed processors, continued strong execution of Nvidia, AMD's resurrection, and hundreds of well-funded startups all focusing on bringing high-performance machine learning processors to the data center, adopting machine learning models will become more affordable in the coming years.
Esoterica's statements are not an endorsement of any company or a recommendation to buy, sell or hold any security. For full disclosures, click here.