After three years of the pandemic, the Qualcomm Snapdragon Summit is finally back in Maui with a series of new processors. In the past three years, Snapdragon has never missed a single huge improvement in smart terminals, and has played the "strongest assist" for smart terminal devices several times to bring users a flagship experience.
Now that the era of generative AI has arrived, how does the new generation of Snapdragon 8 interpret the new benchmark flagship experience?
The benchmark experience of mobile flagships is gradually deeply bound to the Snapdragon 8 series
"Snapdragon is changing the way people live, work, and play, and I'm excited to share that Snapdragon has become the core of nearly 3 billion devices worldwide. At the keynote speech on the first day of the 2023 Snapdragon Summit, Mo Kedong, senior vice president and CMO of Qualcomm, said.
This change is being embraced by more consumers. According to data released by Qualcomm, consumers are more than 9 times more likely to buy Snapdragon terminals than competing products, and smartphone users are 10 times more likely to recommend Snapdragon than other competing products. When it comes to buying a new smartphone, you're three times more willing to buy a smartphone with Snapdragon than other brands. In addition, consumers are willing to pay a 16% premium for Snapdragon terminals, up from 12% last year.
Consumer recognition of the Snapdragon brand is increasing year by year, in addition to the fruitful results brought by Qualcomm's continuous efforts to build the brand value of Snapdragon, it is more about the tangible leading experience brought by the Snapdragon mobile platform to consumers.
Taking the release of the second-generation Snapdragon 8 in the past year as an example, this generation of flagship mobile platform not only adopts the industry-leading TSMC 4nm process, Kryo CPU performance is increased by 35%, energy efficiency is increased by 40%, Adreno 740 GPU performance is increased by 25%, energy efficiency is increased by 45%, and four key upgrades have been made in AI.
These include adding a dedicated power supply system to the Hexagon processor to further improve the energy efficiency of chip processing, improving memory read/write to minimize the reading and writing of external data and reducing energy consumption, and splitting a single neural network into tens of thousands of microslices through micro-slice inference, and then distributing them to tensors, scalars, and vector accelerators to improve the inference speed of the entire neural network. In addition, it is the first to support the INT4 AI precision format.
With the release of the second-generation Snapdragon 8, mobile phone brands such as Xiaomi, vivo, OPPO, moto, and Meizu have all become the first customers to carry them, and they have brought outstanding performance in many aspects, including energy-efficient processing, device-side intelligence, ultra-fast and stable connectivity, and top-of-the-line video, audio, and gaming features.
Moreover, since the second-generation Snapdragon 8, Qualcomm has begun to try to promote generative AI to the device-side experience, and took the lead in supporting the deployment of Stable Diffusion on the device side.
In the face of the advent of the era of generative AI, can Snapdragon bring new surprises and continue to interpret the "flagship benchmark"?
The third-generation Snapdragon 8 has been upgraded to set a new benchmark in the era of generative AI
According to the relevant parameters announced at the Snapdragon Summit, every subsystem of the third-generation Snapdragon 8 mobile platform has been improved, with CPU performance up to 30% higher and GPU performance improved by 25% compared to the previous generation platform. It has also been upgraded in terms of AI and imaging.
"Snapdragon 8 Gen 3 features the most powerful device-side intelligence we've integrated into a smartphone platform to date, enabling breakthrough AI experiences that truly meet the needs of users," said Alex Katuzan, senior vice president and general manager of Qualcomm's mobile, computing and XR businesses, in his speech at the summit.
So, what are the specific aspects of the breakthrough AI experience brought by the third-generation Snapdragon 8, and how is it achieved?
Lekha Motiwala, Director of Product Management at Qualcomm Technologies and Product Owner of the 3rd Gen Snapdragon 8, said that Qualcomm believes that AI and generative AI will undoubtedly be the next disruptive technologies, so when it launched the development of the 3rd Gen Snapdragon 8 mobile platform three years ago, it has put on-the-go intelligence at the top of the list, so that consumers can harness the power of on-device AI and more possibilities to enjoy new use cases.
On the one hand, the enhanced Qualcomm Hexagon NPU makes generative AI inference faster. "In fact, this is our most advanced Hexagon to date, with 98% faster inference and 40% better performance per watt than its predecessor," added Lekha Motiwala.
On the other hand, another key component of Qualcomm's AI Engine, the Qualcomm Sensor Hub, is capable of improving AI performance by 3.5 times and improving memory by 30%. AI assistants connected to large language models (LLMs) work in tandem with Qualcomm Sensor Hubs to provide contextually personalized responses using information such as location and activity data.
In terms of generative AI, the third-generation Snapdragon 8 is capable of supporting models with up to 10 billion parameters, and can run large language models at a rate of 20 tokens per second, maintaining a leading position on mobile devices.
To achieve this speed, Qualcomm quantifies the huge AI model, then compresses it using Qualcomm's own AI software stack tools and loads it into memory.
The chatbot experience is much more intuitive. User speech can pass through Whisper, a Transformer-based automatic speech recognition model, partially running on a Qualcomm sensor hub, and then text-to-text inference will be performed by the 7 billion parameter Llama 2 base model, which has been compressed and quantized.
In order to make the model run smoothly on the terminal, Qualcomm's speculative decoding technology is also used in edge terminals for the first time.
This technique uses a smaller draft model that first performs a lot of inference on the CPU, then quickly generates 3 speculative tokens and notes the model processing tokens, a good draft model predicts the next token with a higher acceptance rate, thus doubling the token generation speed while maintaining accuracy. Finally, the voice generated by the AI assistant is generated by a text-to-speech AI model running on the CPU. Finally, Qualcomm's 5G technology is used to receive information from Skyscanner's cloud servers.
Equipped with these new technologies, the AI capabilities of mobile device manufacturers have also been greatly improved.
Honor CEO Zhao Ming said at the Snapdragon Summit that the Honor Magic6 will be equipped with the third-generation Snapdragon 8 mobile platform and become a flagship mobile phone equipped with an AI device-side model.
Based on the large model of the terminal, HONOR's voice assistant YOYO can also create themed videos based on short prompts provided by users, retrieve raw materials with the same theme from the user's local media library, and change the background music or template through dialogues.
Spread all over smart terminals to create a powerful flagship experience platform
Qualcomm keeps up with the action of the generative AI era, which is not only reflected in the improvement of the third-generation Snapdragon 8, but also in the continuous creation of a Snapdragon product ecosystem across different devices.
At the Snapdragon Summit, Qualcomm announced the launch of Snapdragon Seamless, which enables devices to easily discover each other and connect and collaborate. Based on Qualcomm's low-power, low-latency Bluetooth and Wi-Fi technologies, as well as Always-Aware enabled by Qualcomm's sensor hubs, users are able to create self-organizing intelligent networks that connect smartphones, laptops, tablets, earbuds, smartwatches, and XR devices.
At the same time, Qualcomm also launched the Snapdragon X Elite platform, which opens a new era of intelligent computing with disruptive processing power and AI performance, and can support generative AI models running more than 13 billion parameters on the device side.
There's also the first-generation Qualcomm S7 Pro audio platform, ushering in a new era of audio innovation with device-side AI, revolutionary ultra-low-power Wi-Fi connectivity, and support for Extended Personal Area Network (XPAN) technology.
As we can see, the Qualcomm Snapdragon 8 series is no longer an independent entity, but is closely integrated with PCs, headphones and other smart terminals to jointly create a flagship-level experience and interpret a new flagship benchmark in an ecological way.