Just when we thought the smartphone applications processor (AP) supplier base is shrinking, Google has entered the market with its new Tensor AP. Semiconductors are not unfamiliar to Google. The company designs various semiconductor components, including Tensor Processing Unit (TPU) to accelerate inference and training tasks in datacenters and Video Coding Unit (VCU) to handle YouTube traffic.
In addition to datacenter chips, Google designs a variety of client-centric chips, including Pixel Visual Core, Pixel Neural Core, Soli Radar chip and Titan M security chip for smartphones and EDGE TPU for IoT devices.
Unsurprisingly, Google entered the smartphone AP market with its Tensor AP, which will appear in Pixel 6 and Pixel 6 Pro devices in Q4 2021. Google did not reveal Tensor AP's full specs but indicated that AI and ML would be the key focus for the company. It is improbable that the chip's 5G baseband technology is Google's in-house. While not confirmed officially yet, Google seems to be sourcing 5G modem technology and foundry service from Samsung. Most likely, the Tensor AP integrates the 5G modem. Integrated chips offer multiples advantages including lower power consumption, lower costs and lower circuit board area.
The smartphone AP market returned to growth in 2020 on the strength of 5G and will continue the momentum in 2021 and 2022. Qualcomm led the smartphone AP market with a 40 percent revenue share, followed by MediaTek with 26 percent and Apple with 20 percent in Q1 2021.
What does Google Tensor mean for Google's existing AP and modem supplier Qualcomm? Strategy Analytics estimates that Google contributed 0.6 percent and 0.8 percent of Qualcomm's total smartphone AP units and revenue, respectively, in Q1 2021. In addition to APs, Google sources a variety of chips from Qualcomm, including RF Front-end(RFFE) chips (mmWave module, 4G/5G power amplifier, antenna tuner and envelop tracker, among others).
- In total, we estimate that Google's pixel smartphone lineup represented a slightly less than $150 million opportunity for Qualcomm in 2020.
- We think Qualcomm will retain RFFE and IP licensing revenue in future Pixel devices, limiting the revenue loss from Google Tensor.
Qualcomm has seen increased competition from companies that design their chips in-house. We estimate that vertical vendors such as Apple, HiSilicon/Huawei and Samsung nearly doubled their share to 37 percent in 2020 compared to 19 percent in 2014. The in-house chip efforts already cut into Qualcomm's core smartphone chip business, but Qualcomm offset the loss nicely with investments in RFFE, automotive and IoT businesses. Aside from Google competition, we note that Qualcomm's customers Apple, Xiaomi, Oppo and Vivo are also developing their own 5G modems.
What does this mean for Samsung? While Samsung's 5G modem technology is not yet confirmed in the Tensor AP, we think Samsung is the most likely supplier. Over the last six years, we urged Samsung to expand its cellular chip customer base beyond Samsung Mobile. The company has had some success with Vivo lately with 5G APs. In light of HiSilicon's forced exit, we think Samsung has an excellent opportunity to expand its cellular chip business to emerge as a competitor to Qualcomm and MediaTek.
What does this mean for Google? We think scale, in-house 5G modem technology and accelerated product roadmap are critical ingredients to success in the smartphone chip space. Google is still a sub-scale player in smartphones and the company has stepped up its efforts to gain scale with mid-range Pixel devices in recent quarters. Google doesn't yet have an in-house 5G modem and we urge Google to acquire the modem technology. Even Apple, who passed on multiple opportunities to purchase a modem company earlier, finally made a move and acquired Intel's modem business in 2019.
While AI/ML will be the focus area for Google, we note that over 70 percent of smartphone APs shipped in Q1 2021 already feature an on-device AI engine. In addition, the current smartphone AP AI performance ranges from 2 TOPs to 32 TOPs. In addition to smartphones, Google can expand its Tensor AP to various applications including, Chromebooks, tablets, AR/VR and automotive.
Despite being constrained by computing power, power consumption, memory footprint and network bandwidth, smartphones now feature AI engines with up to 32 TOPs (tera operations per second) of performance. In addition, smartphones currently support various AI frameworks such as TensorFlow, TensorFlow Lite, Caffe, Caffe 2, ONNX (Open Neural Network Exchange), Android Neural Network API (ANN API) and Apple CoreML/ML2.
Machine Learning (ML) is a subset of AI. ML uses algorithms to learn and improve over time with the data it gets exposed to. While the current smartphone AI/ML activity is centerd around inference, we think future smartphone APs will feature training engines. Google, for example, enabled a conversational machine learning model on smartwatches without internet connectivity.
- Smartphone chip vendors opt for lower precision (16-bit, 12-bit, 8-bit or lower) to address inference tasks on the device. Lower precision can potentially reduce the neural network model accuracy, but research has shown that training and inference can be performed with lower precision without sacrificing accuracy. The lower precision inference has multiple advantages, including lower memory usage, lower silicon area and lower power consumption than higher precision models. In addition, vendors such as Qualcomm include 32-bit support to support complex ML applications.
We note that NVIDIA-Arm is also focused on proliferating AI/ML in edge devices such as smartphones. Google's entry further heats up the smartphone AI chip market.
In summary, we think Google's entry into smartphones could further AI innovation, but we must note that scale and modem technology is crucial for Google to win this space. Therefore, we urge Google to buy or develop 5G modem technology in-house.
Sravan Kundojjala
@SKundojjala