The annual COMPUTEX has come to an end. The theme of this year is “AI Connects, Creates the Future Together”. Compared with the “dismal” performance of previous years, this exhibition is full of highlights due to the popularity of AI technology, especially in the AI PC field which has been very popular in the past two years. Intel, AMD and Qualcomm have successively “shown their swords” and shared their latest products, layouts and progress. Today, I will take you to review their new developments in the AI PC field at this exhibition.
Intel disrupts again
Advertisement
Last November, Intel released the Core Ultra processor (Meteor Lake) designed for AI PCs. It not only has the new Intel 4-process technology and 3D packaging hybrid architecture, but also has new performance cores and energy efficiency cores and Arc GPU. It also integrates an NPU designed specifically for AI. Combined with the CPU and GPU, it can more flexibly respond to AI computing power requirements in different scenarios, bringing what Intel calls the biggest processor architecture change in 40 years.
At this year's COMPUTEX 2024, Intel announced details of the Lunar Lake architecture for the next-generation AI PC, including modular structure, packaging process, new performance cores and energy efficiency cores, thread scheduling, GPU, NPU and connectivity, etc. It can be said that it has once again subverted the previous ones.
Let's talk about them one by one. First, the modular structure and packaging. Lunar Lake still continues the separate module design of Meteor Lake, but it is simplified into two parts: the compute module (Compute tile) and the platform controller module (Platform Controller tile), and uses 3D Foveros packaging technology.
Advertisement
More importantly, Intel integrated the memory into the package for the first time, calling it “package-level memory”. The capacity of the two memory chips is up to 32GB, and it supports up to LPDDR5X 8533.
Packaging inside the SoC shortens the memory routing and can reduce physical power consumption by 40%. Of course, this also means that Lunar Lake terminal devices will not be equipped with memory slots and subsequent expansion will not be possible.
Taking into account the memory usage of current large-scale applications and generative AI, it is estimated that the next generation of terminals will mostly have 32GB of memory.
Next are the new performance cores and energy efficiency cores. Lunar Lake uses the new Lion Cove performance cores, which have an average IPC improvement of about 14% compared to the Meteor Lake performance cores; and even more amazing is the Skymont energy efficiency core, which has an average floating-point IPC performance improvement of 68% compared to the LP energy efficiency cores in Meteor Lake.
Intel said that four energy-efficient cores form a cluster. Compared with the two LP energy-efficient cores in Meteor Lake, their power consumption is only 1/3 of theirs at the same performance, while their performance is 2.9 times that of theirs at the same power consumption, and their peak performance is 4 times that of theirs.
This time, Lunar Lake has a maximum of 4 performance cores and 4 energy efficiency cores. There is no separate LP energy efficiency core. Instead, this part is integrated into the CPU computing cluster, and the performance core has cancelled hyper-threading. Yes, there is no hyper-threading, which means that this generation of Lunar Lake has a maximum of 8 cores and 8 threads, which is indeed a bit surprising. From this we can see that Intel did not choose to pile up the number of cores this generation, but focused on efficiency.
In addition, Intel has also designed a new hardware thread scheduler for the new architecture, which uses a dynamic scheduling strategy. When there is a suitable load, it will give priority to a single energy-efficient core. When the load is larger and more performance is required, other energy-efficient cores will be called first. When the performance demand increases further, it will turn to the performance core to obtain better energy efficiency performance.
Then there are the GPU and NPU. The former has been upgraded to the second-generation Xe2 microarchitecture and has 8 second-generation Xe cores. Its performance is about 1.5 times that of the previous generation and it can provide up to 67TOPS of computing power. The latter is called NPU 4, the fourth generation. It has 6 neural computing engines, which is 3 times that of the previous generation, as well as 12 enhanced SHAVE DSPs, energy-efficiently optimized MAC arrays, 2 times the bandwidth, and supports native activation functions and data conversion, etc., with a computing power of up to 48TOPS.
At the same power consumption, NPU4 has twice the performance of NPU3, and its peak performance is four times that of NPU 3.
Regarding the overall computing power of the Lunar Lake platform, the CPU is 5TOPS, the GPU is 67TOPS, and the NPU is 48TOPS, with a total computing power of 120TOPS. The three major engines perform their respective duties and respond efficiently and flexibly to the demands of different loads.
Judging from the information released so far, Lunar Lake can be said to have once again achieved disruptive changes compared to its previous generation. It has to be said that Intel has indeed invested a lot of energy in the field of AI PC.
I think the most noteworthy aspect is energy efficiency. Judging from the official data so far, the improvement is very large, but data is just data after all, and the actual improvement still has to be achieved at the product level. There is also a significant improvement in AI computing power, which has increased several times compared to the previous generation. Stronger computing power should bring obvious improvement to the experience of related AI applications. Of course, this also needs to be verified after the actual product is launched.
Terminal devices equipped with Lunar Lake processors should be available later in the second half of the year. It will be worth looking forward to whether disruptive changes can bring about disruptive experiences.
AMD is moving forward
As previously revealed, AMD renamed the new mobile processor for the next-generation AI PC as the Ryzen AI 300 series. Its CPU uses the latest Zen 5 architecture, the GPU is an upgraded version of the RDNA3.5 architecture, and the NPU is the new XDNA2 architecture. The three major computing units of CPU, GPU, and NPU have been comprehensively upgraded, but it still continues the 4nm process.
AMD not only changed the naming of its products this time, but also added AI directly into the model name, which shows their emphasis on AI. The reason why they started directly with the 300 series is because this is the third generation of AI PC processors.
The first batch only has two models, Ryzen AI 9 HX 370 and Ryzen AI 9 365, both of which are positioned in the high-end market. It should be noted that the meaning of HX here is different from the Ryzen 9 7945HX for high-end gaming laptops, and should represent a more high-end meaning.
Among them, the Ryzen AI 9 HX 370 has 12 cores and 24 threads, a maximum acceleration frequency of 5.1GHz, 24MB of L3 cache, a Radeon 890M GPU, 16 CU units, and the new XDNA2 architecture's NPU computing power reaches 50TOPS; the other Ryzen AI 9 365 has 10 cores and 20 threads, a maximum acceleration frequency of 5.0GHz, 24MB of L3 cache, a Radeon 880M GPU, 12 CU units, and the NPU part remains consistent with the above.
It can be seen that the Ryzen AI 300 series not only has an architectural upgrade, but the CPU part also breaks the 8-core 16-thread system that has been in place for many years. Yes, AMD has started to pile up the number of cores, and they are all large cores, which is quite interesting. The GPU also provides more CU units, and the NPU has achieved more than 3 times the computing power compared to the previous generation.
According to official data from AMD, the XDNA2 architecture NPU not only has a 5-fold improvement in computing performance (Ryzen AI 9 HX 370 vs. Ryzen 9 8940HS in the response speed test of the Llama 2 7 billion parameter large model from startup to obtaining the first token), but also has doubled its energy efficiency and multi-task parallel capabilities. The improvement of end-side AI computing power can effectively improve the local AI experience, give full play to the advantages of AI in low latency, security, personalization, etc., and provide sufficient computing power foundation for the efficient and smooth operation of related AI applications.
Compared with competing products from Qualcomm, Apple, and Intel, the Ryzen AI 9 HX 370 has achieved a clear lead in productivity. In terms of gaming, it leads the Core Ultra 9 185H by an average of 36%. However, the Intel products it is comparing with are all previous generation products, so it is not very meaningful.
Of course, in addition to the hardware upgrade, AMD once again emphasized its layout in ecological applications. By 2024, there will be more than 150 ISV partner manufacturers, including giants such as Adobe, Microsoft, Topaz Labs, as well as domestic companies such as Baichuan Intelligence, DingTalk, Wuwen Xinqiong, and Youdao. After all, no matter how powerful the AI computing power is, it will be meaningless if it is not implemented in applications for users.
Compared with the “toothpaste squeezing” of the previous generation, this time the Ryzen AI 300 series can be said to be a big step forward. The CPU and GPU have more cores, and the NPU computing power has become the strongest at present. Although it has not subverted itself again like Intel, there are still many highlights. Moreover, this generation even added “AI” to the model, which shows how much AMD attaches importance to AI PC.
Of course, there is actually no obvious gap between the two companies in terms of computing power. In comparison, the competition points are more about the actual experience of AI applications and the construction of the software ecosystem.
Qualcomm's ambitious
Unlike Intel and AMD, which focus on future new products, Qualcomm did not release any new products at this year's COMPUTEX 2024. Instead, it focused on how PCs equipped with the Snapdragon X series are empowering the transformation of the PC industry. The theme of “Reshaping the PC” undoubtedly shows Qualcomm's ambitions in the fields of PC and artificial intelligence. They will provide hardware computing power, lay out the software ecosystem, and empower developers.
At the Snapdragon Summit at the end of 2023, Qualcomm released the flagship Snapdragon X Elite processor, and in April this year, it brought the sub-flagship Snapdragon X Plus. The outstanding performance release, excellent performance per watt and lower energy consumption under the same performance are the advantages of the Snapdragon platform. For thin and light notebook devices, the leading features of these features are of great practical significance.
At the same time, NPU, as a key differentiating factor, enables PCs equipped with the Snapdragon X series to have excellent performance. Shifting AI workloads from the CPU and GPU to the NPU can significantly improve performance while reducing power consumption.
Compared with Apple's M3, the performance per watt of the Snapdragon X Elite NPU is 2.6 times that of Apple's M3 and 5.4 times that of the Core Ultra 7. In the words of Qualcomm President and CEO Amon: Qualcomm Technologies is reshaping the performance leadership of the Windows PC ecosystem.
The advantages of Qualcomm NPU can be reflected in many different application fields. With up to 45TOPS of AI computing power, almost all application fields of PC can use NPU to create a better user experience. In the field of software development and programming, software developers can use the 45TOPS NPU computing power on the terminal side to generate code in real time; these computing powers can help users realize rich applications such as image quality enhancement, AI filters, and AI noise reduction.
Snapdragon X Elite also supports large language models with 7 billion parameters running at a speed of 30 tokens per second on the terminal side, and many cases have been put into practical use. Qualcomm also demonstrated the application of AI in music production, video editing and gaming experience, as well as commercial AI model cooperation with partners such as Deloitte, highlighting the potential of AI PC in improving productivity and innovation.
In addition to striving to provide users with a better experience, Qualcomm has launched a unified software stack that is cross-platform, cross-terminal, and cross-operating system for developers – the Qualcomm AI Stack. It supports all current mainstream training frameworks and execution environments, and provides developers with optimization interfaces at different levels and levels, as well as a complete compilation tool chain, allowing developers to more efficiently complete model development, optimization, and deployment on the Snapdragon platform.
Previously, Qualcomm also launched the Qualcomm AI Hub model library, providing developers with more than 100 pre-optimized AI models, including generative AI models including language, text and images, as well as traditional AI models such as image recognition, image segmentation, natural language understanding, and natural language processing, to accelerate the development of AI applications.
Qualcomm not only provides underlying computing power for devices through the Snapdragon X series chips, but also works with ecosystem partners to create applications and provide developers with convenient development tools. It also has close cooperation with Microsoft. As the middle layer connecting hardware computing power and AI applications, the operating system also plays a key role. It can be said that Qualcomm has made a full-scale layout at the software and hardware levels.
Although Qualcomm's products are a bit “late” compared to Intel and AMD's AI PCs, from another perspective, they have been prepared for a longer time and should be more fully prepared. Currently, many OEM manufacturers' new products are already on sale. We can wait and see what their actual performance is and whether they can really bring about a “reshaping PC” experience.
Conclusion
Judging from COMPUTEX 2024, AI PC has obviously become a “battlefield”, and the three major “players” as the underlying computing power providers, Intel, AMD and Qualcomm, have all shown their talents.
Both Intel and AMD announced their next-generation processors. In comparison, AMD's Ryzen AI 300 series has achieved comprehensive improvements in the three major computing units of CPU, GPU, and NPU, so I used the word “big strides forward” to describe it, but the overall product logic has not changed very much.
The changes in Intel's Lunar Lake can be more appropriately described as subversive. It not only focuses on improving AI computing power and energy efficiency, but also overturns many features and functions of previous products. To be honest, this is a bit beyond everyone's expectations. After all, the previous generation has just been greatly changed, but this generation did not choose to steadily improve, but to subvert again. It can only be said that Intel has indeed put enough effort into AI PC.
As for Qualcomm, although there is no new processor launched this time, it undoubtedly demonstrated its great ambition to become an industry disruptor through AI PC. At the same time, AI PC equipped with Snapdragon processor has been on the shelves, but how it performs still needs to be verified by consumers and the market.
But no matter what, competition in the AI PC market will definitely become more intense in the second half of the year. This competition is not only between underlying computing power providers such as Intel, AMD and Qualcomm, but the competition among OEM manufacturers will also be very fierce. Using chips with the same computing power, how to win the favor of consumers through differentiated experience, better ease of use, richer ecological applications, smarter interactive experience, etc., all require a lot of effort.
