Abstract:
Qualcomm recently announced further details of the AI processing capabilities of Snapdragon 8 Elite Gen 6 Pro, the most noteworthy of which is the upgrade of the new Hexagon NPU. Different from purely pursuing the growth of AI computing power in the past, Qualcomm this time pays more attention to the efficiency of AI model operation, memory access and power consumption control, hoping to enable the next generation of flagship mobile phones to run larger-scale generative AI models and AI agents locally.
Qualcomm said that as generative AI gradually becomes an important feature of smartphones, the AI models themselves continue to become larger and more complex. For mobile devices, the real bottleneck is not just computing power, but how to efficiently process these models under limited power consumption and memory bandwidth. Therefore, the NPU of Snapdragon 8 Elite Gen 6 Pro has been redesigned in many aspects.
One of the important upgrades is the new Element Accelerator. This dedicated hardware acceleration unit is optimized for specific operations in AI computing, which can improve the efficiency of the NPU when executing related workloads. Compared with simply increasing the peak computing power of the NPU, this specialized design allows the chip to complete specific AI tasks with fewer resources, thereby reducing energy consumption in the computing process.
The new generation of Hexagon NPU also gains larger shared memory. Qualcomm said that its Large Shared Memory capacity has increased by 50% compared to the previous generation. This part of high-speed shared memory allows the NPU to more conveniently save the data required during the running of the AI model, thereby reducing the need for frequent access to system memory.
This change is especially important for large language models. When running generative AI, models need to constantly process large amounts of contextual data, some of which must be persisted during inference. If this data is frequently transferred back and forth between the NPU and system memory, it will increase latency and consume more power. By adding high-speed shared memory that can be used directly inside the NPU, Qualcomm hopes to reduce this data handling as much as possible.
One of the objects that has benefited significantly is KV Cache. As users have longer and longer continuous conversations with AI assistants, models need to save more and more contextual information. The KV Cache size will also increase accordingly, and it often becomes an important memory burden when running large models locally. Larger shared memory allows more relevant data to stay closer to the computing unit, thereby improving AI inference efficiency.
Qualcomm believes that this capability is particularly important for AI agents. Compared with traditional question-and-answer AI, AI agents need to complete multiple steps continuously, such as understanding user needs, formulating execution plans, calling different tools, analyzing the returned results, and then proceed to the next step. Throughout the process, model state and context need to be continuously saved, so reducing memory access can help reduce latency and allow complex AI tasks to run more smoothly.
Qualcomm’s improvements to the NPU also mean that when mobile phones run AI in the future, they do not necessarily need to simply rely on increasing system RAM to solve the problem. As local AI models become larger and larger, mobile phone manufacturers have continued to increase the memory capacity of flagship products in recent years. However, increasing RAM will not only increase hardware costs, but also increase power consumption and pressure on memory supply. Qualcomm is trying to alleviate this problem at the architectural level by putting more data into high-speed shared memory inside the chip.
In addition to the NPU, the CPU and GPU of the Snapdragon 8 Elite Gen 6 Pro have also been significantly upgraded. Qualcomm has previously revealed that the new generation flagship platform will have a CPU frequency of up to 5GHz. If this specification is finally implemented according to the currently announced information, it will become a very symbolic node in mobile processors, because the highest frequency of smartphone CPUs will officially enter the 5GHz era.
Qualcomm also introduced FlexCache technology for CPUs. The core idea is also to reduce the processor's dependence on external system memory and make more flexible use of the chip's internal cache so that the CPU can obtain data more efficiently. For modern mobile processors, data access efficiency between cache and memory has become increasingly important, so this design can improve the balance between performance and energy efficiency to a certain extent.
GPUs have also seen significant architectural changes. Qualcomm describes the new generation of Adreno GPU as a very important architectural upgrade that not only improves traditional graphics performance, but also further enhances the GPU's ability to perform AI tasks.
One of the important changes is the addition of Adreno Matrix Cores, which are cores dedicated to matrix calculations. Matrix operations are an important type of calculation in modern AI models, so dedicated matrix cores can help GPUs perform AI-related work more efficiently.
This means that future Snapdragon flagship platforms will not hand over all AI calculations to the NPU. Depending on the specific tasks, CPU, GPU and NPU can each undertake different types of work, thus forming a more obvious heterogeneous computing system. For games, the GPU can perform part of the AI calculations while completing traditional graphics rendering; while for workloads such as generative AI, the NPU can play a more important role.
Qualcomm also launched Adreno Neural Fusion technology to further strengthen the combination between AI and graphics rendering. This technology can use AI algorithms to improve the visual effects of game screens and help achieve more advanced graphics processing effects while controlling power consumption.
This design is particularly suitable for AI graphics technologies that have developed rapidly in recent years, such as super-resolution and frame generation. The resolution and picture complexity of mobile games continue to increase. If all pictures rely on traditional methods for native rendering, the GPU burden and power consumption will increase rapidly. Using AI to generate part of the picture information can achieve higher final picture quality at lower rendering costs.
From the perspective of the overall architecture, the AI upgrade of Snapdragon 8 Elite Gen 6 Pro is not simply to make the NPU faster, but to try to redesign the data processing method inside the entire SoC. The NPU has larger shared memory and dedicated accelerators. The CPU reduces dependence on system memory through FlexCache, while the GPU adds specialized matrix computing capabilities and participates in graphics processing through AI technology.
This change reflects that mobile chips are entering a new stage. In the past, when measuring the performance of flagship SoCs, people mainly focused on CPU and GPU running scores, but now AI workload has become an important factor in determining the actual experience of the chip. For the next generation of mobile phones, what is really important is not only how high the peak computing power of the chip can be, but also whether it can run AI models for a long time under limited power consumption, and whether it can effectively control the movement of data between different computing units and memories.
Qualcomm's special emphasis on shared memory and reducing memory access also shows that chip competition in the AI era is gradually shifting from a simple "computing power war" to a "data transfer efficiency war." For large AI models, the calculation itself is important, but how to send the model parameters, context and intermediate results to the correct computing unit in a timely manner will also directly affect the final performance and energy consumption.
Therefore, what is really noteworthy about the Snapdragon 8 Elite Gen 6 Pro may not be how much individual AI performance has been improved, but that Qualcomm is trying to enable mobile phones to handle increasingly complex AI tasks locally through closer collaboration between NPU, CPU, GPU, cache and shared memory.
As AI assistants, AI agents, generative applications, and AI game functions gradually become an important part of flagship mobile phones, this architectural idea may also become an important direction for the development of mobile SoCs in the future. Compared with continuously increasing system RAM, digesting the ever-expanding AI models through more efficient data processing and caching mechanisms within the chip is expected to achieve a more reasonable balance between performance, power consumption and cost.
Whether the Snapdragon 8 Elite Gen 6 Pro can ultimately achieve these technical goals will need to be verified through complete testing after the actual device is launched. However, judging from the information released so far, the focus of upgrading Qualcomm's flagship platform of this generation has been very clear: not only to make AI calculations faster, but also to make AI calculations more efficient, and to minimize the memory and power consumption costs of mobile phones in order to run local AI.
Comments