October 2021. OPPO has just unveiled MariSilicon X, its first in-house imaging chip, and the announcement is more significant than the framing it is receiving in Western tech coverage. The MariSilicon X is not a marketing chip. It is not a co-branded sensor. It is a fully custom NPU (Neural Processing Unit) designed from the ground up for imaging workloads, built by OPPO’s chip design team, manufactured at TSMC. The specifications: 18 TOPS of AI computing at 11.6 TOPS per watt, 20-bit dynamic range processing, 8.5GB/s dedicated DDR memory bandwidth for the imaging pipeline. OPPO is saying it enables 4K AI Night Video, Ultra HDR capture, and real-time RAW processing. The chip ships on the Find X5 series in 2022.
For context: Google announced its Tensor chip in August 2021. Apple has been building custom silicon for imaging and neural processing since the A11 Bionic in 2017. Xiaomi’s Surge C1 ISP co-processor arrived in the Mi 11 Ultra in March 2021. OPPO is the fourth company to develop in-house silicon for mobile imaging, and the first to build a dedicated NPU rather than an ISP. The distinction matters: an ISP handles raw sensor data conversion. An NPU handles the machine learning inference that determines what the image looks like after processing. OPPO has decided that the limiting factor in its imaging quality is the AI layer, not the sensor conversion layer, and has built silicon specifically for the AI layer.
Why build your own chip
The standard answer is “control.” If you control your imaging NPU, you control the image quality outcomes rather than being constrained by the ISP capabilities of whatever Qualcomm or MediaTek ships you. This is Apple’s argument for building its own silicon across the full stack, and it is a compelling argument when your imaging quality is a primary competitive advantage. OPPO’s imaging has been competitive-but-not-leading relative to Apple and Google throughout the Find X2 era. The MariSilicon X is an attempt to move from competitive to leading by owning the processing layer that determines the quality ceiling.
The honest caveat: new silicon takes time to reach its potential. Google’s Tensor, which launched in August 2021, is not yet operating at the level its architecture implies it can reach. The ISP algorithms, the denoise models, the computational photography pipeline: all of these take multiple software update cycles to optimize for a new chip’s capabilities. MariSilicon X launching on the Find X5 in early 2022 will be the beginning of OPPO’s in-house imaging story, not the complete chapter. Whether the chip proves its value over six to twelve months of software updates is the test that will determine whether OPPO’s chip investment was the right strategic bet. The Snapdragon 888 inside the same Find X5 already has three 18-bit ISPs with 26 TOPS of AI capability. The MariSilicon X needs to do something meaningfully different to justify the years of engineering investment behind it.
I am watching this with genuine interest. If OPPO’s chip team has built what the specifications imply they have built, the Find X5 will have camera output that evolves visibly over its first year of software updates. That evolution would be the proof. We will know in 2022.
Part of OPPO: The Complete Story, my full history of OPPO from DVD players to the BBK empire.
Related reading: how Xiaomi’s own Surge ISP ambitions fit the same in-house-silicon push.