Regarding the new generation flagship processor announced at the Snapdragon Tech Summit 2023Snapdrgaon 8 Gen 3Qualcomm later explained more details through an in-depth Q&A session after the meeting, including that the Snapdragon 8 Gen 3 processor maintains the same process while significantly improving performance. The company even directly revealed in the Q&A that the next processor product will be called Snapdragon 8 Gen 4.
While it's unclear whether Chris Patrick, Qualcomm's Vice President and General Manager of Mobile Handheld Devices, directly referred to the Snapdragon 8 Gen 4 name in response to media questions, he did acknowledge that the company's processor products, which will utilize its fully proprietary Oryon CPU architecture next year, may see increased costs, but this will also result in higher computing performance.
As for the newly launched Snapdragon 8 Gen 3, while maintaining the use of TSMC's 4nm process, it can achieve a significant performance improvement. Chris Patrick said that the adjustments mainly focus on the core configuration and operating clock. For example, this time it adopts a special "1+5+2" core configuration, abandoning the previous design of 4 energy-saving cores and retaining only 2 minimum necessary energy-saving core configurations, allowing the processor to operate at a higher efficiency.
If you look closely at the die area design of Snapdragon 8 Gen 3, you will find that it is actually relatively increased. However, when combined with artificial intelligence technology, the processor can produce higher computing performance with relatively lower power consumption.
Chris Patrick also revealed that after the official launch of the new flagship processor equipped with the Oryon CPU design in 2024, application products targeting the mainstream and entry-level markets will indeed be launched, and the artificial intelligence design will also be adjusted accordingly to correspond to different product positioning, but it is claimed that a certain device-side artificial intelligence computing capability will still be retained.
As for the Snapdragon 8 Gen 3, it is claimed that it can run large natural language models with more than 100 billion parameters, but it can also be adjusted to 70 billion parameters in a more flexible way, and through INT4 precision optimization, it can run smoothly on mobile phones with a minimum of 8GB of memory.
If 8 billion sets of parameters are executed with INT70 precision, approximately 7GB of memory capacity is required. However, if it is converted to INT4 precision operation mode, only 3.5GB of memory capacity is required for smooth execution.
Given that the Snapdragon 8 Gen 3 processor can support up to 24GB of memory, it can successfully execute over 4 billion sets of parameter calculations with INT100 precision, and is compatible with various device-side automatic generative artificial intelligence application functions. For example, the image edge extension function demonstrated this time can generate possible corresponding scenes through automatic generative artificial intelligence in a short time.
