When the world’s attention was on Open AI and Google, Microsoft quietly dropped a bombshell.
In the early morning of May 21, Beijing time, on the eve of the Microsoft Build Developer Conference, the software giant held a special press conference and proposed a concept called “Copilot+PCs”.
Compared with the “AI PC” that was frequently displayed by the host manufacturers but had a vague concept, Microsoft integrated the newly released GPT-4o at the Windows system level; at the hardware level, the built-in Snapdragon X Elite chip can process generative AI applications locally without relying on cloud computing power.
Although there were rumors that Microsoft would upgrade Copilot, this press conference still exceeded expectations.
First of all, it was unexpected that GPT-4o could be integrated into the terminal product so quickly. Of course, considering the relationship between Open AI and Microsoft, the cooperation between the two parties on end-side AI has probably already begun.
Secondly, the first batch of “Copilot+PCs” products are all equipped with Snapdragon X Elite/X Plus chips, which means that Microsoft has begun to lead the host manufacturers to collectively turn to the ARM camp.
In the past few years, Apple has been eroding the Windows camp in the PC market with its ARM-based M series chips. Now Microsoft has chosen to launch a counterattack against the company in the way that Apple is most familiar with.
At the press conference, Microsoft executives even placed a Macbook Air equipped with an M3 chip next to the demonstration product to show how Copilot+PCs is far ahead of the latter in the image generation process.
Image generation speed comparison, image source: Microsoft
Microsoft is so openly “tit-for-tat”, can Apple still take the trick?
AI PC is finally a little bit clear
Compared with the improvement in processing speed of generative AI applications, what is more commendable at this press conference is that Microsoft demonstrated many innovative applications that have never been seen in the industry.
For example, let your computer “go back in time”.
This feature, named “Recall”, allows Copilot to help users find local files, tasks, or web browsing history.
The principle of this function is that the computer will capture and record the current screen snapshot every few seconds and encrypt and store it on the computer hard disk.
We can imagine a scenario: a few weeks ago, you browsed a web page about car colors, and the only valid information you remember was “I saw a red car.” Under similar vague information, you can search for “red car” and let Copilot call up all web pages with the element of “red car” in the past period of time.
Image source: Microsoft
If there are too many results, you can also use the timeline above to select a specific time interval to further filter the desired results. It is worth mentioning that on this timeline, you can go back to any task you have previously operated, even the videos you have watched.
This is definitely a leap in productivity, but the disadvantages are also obvious.
The first is the issue of privacy and security. If the privacy and security in the past was a dispute between users and manufacturers over cookies, then on Copilot+PCs, this data trace has become a “screen recording” that needs to be performed every few seconds on the user’s computer.
After the press conference, Musk bluntly said, “This is an episode of Black Mirror. This feature must be turned off.”
Image source: @Elon Musk, X
In this regard, Microsoft mentioned many times at the press conference that all recorded information will be stored locally on the computer and will never be uploaded to the cloud, and users can also customize what kind of information the system can record.
However, if all are archived locally, will the massive amount of information retained by the system every day take up too much storage space? Microsoft did not explain this issue on the spot, but said that the “Recall” function will be officially launched later this year.
Another exciting upgrade is that Microsoft showed that with the support of GPT-4o, we finally see what a voice assistant should look like.
In a demo on the spot, the tester opened the game “Minecraft” and called Copilot to ask how to build equipment. Copilot provided suggestions immediately, and even after discovering the “zombie”, it told the player “This is a zombie, you need to get away from it quickly.”
Image source: Microsoft
In fact, if Copilot is regarded as an NPC in the game, similar AI-driven NPCs have been common in recent years. However, it should be noted that in Microsoft’s demonstration scene, Copilot is not a pre-set process, but a real-time call by the player. It needs to process the voice messages sent by the player while monitoring the game screen.
This may bring a disruptive change to the game industry.
In addition, Microsoft has also made substantial upgrades to the “relatively traditional” generative AI scene. On the Surface prototype demonstrated on site, the languages supported by real-time subtitles have been expanded to 40+, and the scenes include video conferencing and streaming content.
Microsoft has also upgraded the diffusion algorithm for its own Paint application. Now users can use simple lines to outline the frame, and the computer will generate the corresponding picture in real time.
Image source: Microsoft
Although it doesn’t sound new at the moment, it should be noted that all these functions are completed with the help of local computing power, and the Snapdragon X Elite, which was previously highly expected by the industry, supports Copilot+PCs.
The PC industry is really going to change.
Two months ago, Intel held a developer conference in Taipei and released a rare definition of AI PC with Microsoft: it must be equipped with NPU, CPU and GPU, support Microsoft’s Copilot, and have a physical Copilot key directly on the keyboard.
The original Windows key has now become the Copoilot key.
Through this statement, we can dig out two key pieces of information: first, in terms of the definition of AI PC, Intel and Microsoft have the same interests, and the Wintel alliance will still exist in the future AI PC era; second, it has become a consensus among all parties to use heterogeneous computing platforms as the hardware foundation of AI PCs.
But the problem is that Intel is not the only one that can play with heterogeneous computing, so at today’s press conference, we saw the combination of “Windows+Qualcomm”.
According to the information released by Microsoft, the first batch of “Copliot+PCs” products will start in mid-June, and host manufacturers including Lenovo, Asus, Dell, HP and other mainstream brands have responded, and all of these products will be equipped with Qualcomm Snapdragon X Elite/X Plus chips.
Image source: Qualcomm
This chip, released last October, is made with TSMC’s 4nm process and is equipped with a new customized Oryon CPU. In the 4+4+4 cluster, the main frequency of the 12 cores can reach 3.4GHz. The NPU equipped with it has a computing power of 45TOPS, and the integrated Adreno GPU can also provide 4.6TOPS computing power.
In comparison, the total computing power of Intel Core Ultra, which is currently commonly installed on AI PCs, is about 34TOPS.
In addition to leading in computing power, the X Elite developed based on the ARM architecture also has an inherent advantage in energy consumption performance over a number of X86 architecture chips.
In fact, Microsoft made the transition from X86 to ARM as early as 2012, and launched the Surface based on the ARM architecture chip that year, but the instruction set used by the ARM architecture at that time was 32-bit, and program compatibility was almost non-existent.
In the V9.2 architecture updated by ARM last year, the designs of the three processor chips, including Cortex-X4, A720 and A520, all turned to 64-bit instruction sets. At the same time, large applications including Photoshop, Zoom, Chrome and others have also launched native versions that support ARM 64-bit.
It should be said that the transition of ARM architecture chips from mobile phones to PCs is an unstoppable trend today.
Of course, considering the strong versatility and flexibility of X86 architecture chips, the purpose of Microsoft’s cooperation with Qualcomm may not be to replace Intel, but to challenge Apple.
For now, at least compared with the M3 chip (the M4 chip has not yet been installed on the Macbook), the combination of Snapdragon X Elite + Windows is significantly stronger than M3 + MacOS.
First of all, in terms of performance, whether it is Geekbench or Cinebench, the Surface Pro equipped with Snapdragon X Elite beats the M3 version of the Macbook in almost every benchmark test.
And Apple’s most proud energy consumption performance, after switching to ARM architecture, the Surface is also far ahead of the Macbook.
Microsoft said that the 15-inch Surface Pro equipped with Snapdragon X Elite has a battery life of 22 hours. In actual tests, compared with the M3 version of Macbook, the former has a 20% longer battery life while the computing speed is increased by 58%.
More importantly, Microsoft has demonstrated the application scenarios and development potential of AI PC through this press conference, while Apple’s actions on AI are still only presented in the form of auxiliary functions, which may not be enough to convince users to pay for it.
Whether it is the marriage between Microsoft and Qualcomm or the active response of various host manufacturers, it can be seen that a closer ecological alliance is taking shape in the Windows camp.
WWDC in a month may be Apple’s last chance.
