Intel recently released the 101.4578 Beta version of its graphics driver, with a primary focus on supporting the Arc series. However, beyond the regular updates, there is a hidden detail.
During the installation process, if you select the typical mode, a new option for the “Compute Improvement Program” (CIP) will appear, which is checked by default.
Like all similar improvement programs, it is used to collect data.
According to Intel’s explanation, to improve the performance, features, and user experience of Arc graphics cards, the CIP program will collect information on users’ computer usage habits, system information, other device information, and websites visited (not including URLs), among other things.
The usage information includes:
- Graphics card frequency
- Duration of driver software use
- System memory usage
- Laptop battery life
- Plug-and-play devices, and more.
System information includes, but is not limited to:
- The device manufacturer
- CPU model
- Memory and display configuration
- Operating system version
- Software versions
- Regional and language settings
- Location and time zone.
Intel also emphasizes that it will never collect sensitive personal information such as a user’s name, address, email, or phone number.
These practices are actually standard for manufacturers, but having the option checked by default can certainly be a bit unsettling.
However, Intel isn’t the most extreme. NVIDIA is the most aggressive; its GeForce graphics drivers have mandatory data collection enabled by default, which cannot be disabled, not even through a custom installation.
AMD provides an option at the end of its Adrenalin graphics driver installation. Regardless of whether you choose the typical, driver-only, or custom installation method, you can opt in or out of data collection.
But regardless, NVIDIA’s cards—both for gaming and acceleration—are dominant right now.
Although the CEO of OpenAI has denied it, the industry believes they are already training the GPT-5 large model, which is expected to be more than 10 times the scale of GPT-4. This means burning through more money, especially as the graphics cards needed for AI training are extremely scarce.
Exactly how many AI graphics cards are being used by the various companies developing large AI models is a closely guarded secret, and accurate data is rarely made public. It is speculated that GPT-4 was trained on 10,000-25,000 A100 GPUs. The number of cards needed for GPT-5 remains a mystery, but Elon Musk estimates it will be between 30,000 and 50,000 H100 GPUs, which are significantly more powerful.
In addition to OpenAI, demand for high-performance AI graphics cards from other companies like Meta, Google, Microsoft, Apple, Tesla, and Musk’s own ventures remains high. The total demand is projected to reach 432,000 H100 GPUs, worth over $15 billion.
And this is just the demand from the US tech sector, as analyzed by netizens abroad. In reality, demand within China is no less significant. Even though Chinese companies can only purchase the special-edition A800 and H800 accelerator cards, this hasn’t stopped them from diving into large AI models, even paying inflated prices to get their hands on AI GPUs.
In the current market, only NVIDIA can satisfy the demand for AI graphics cards. The AI cards from AMD and Intel not only have performance and ecosystem issues but also can’t keep up with supply. AMD’s heavy hitter, the MI300X, won’t ship until the end of the year, with mass availability slated for 2024.
Ultimately, NVIDIA is making a killing once again. The demand for AI graphics cards is expected to continue through the end of 2024, with supply falling short of demand for the next year and a half. The price of an H100 card starts at 250,000 RMB, but with markups, the final cost is hard to predict due to large price fluctuations.