10:59Machine-translated from Azerbaijani by AIRead in Azerbaijani

Traces of US-Restricted Nvidia Chips Found in Chinese AI Models

    An independent research group called “UnemploymentLabs” has found strong evidence that Nvidia's “Blackwell” chips were used to develop Chinese artificial intelligence (AI) models. The US has imposed restrictions on the export of these chips. The study names the “DeepSeek V4,” “Kimi K3.1,” “MiniMax-M3,” and “Ling-3.1-Flash” models. However, it is not known where the chips are located or how the companies gained access to them. Therefore, a violation of export rules has not yet been proven.ABŞ-nin məhdudlaşdırdığı Nvidia çiplərinin izi Çin AI modellərində tapılıb    The study is based on statistical traces left in an AI model's weights when training begins. Weights are numbers that determine how much importance the model assigns to each piece of incoming data. Different chip families have different numbers of computing units known as SMs. “Hopper” chips have 132 such units. The “Blackwell B” series has 148, while the “Blackwell GB” series has 152.    A model's initial weights are randomly generated using “PyTorch,” “CUDA,” and “cuRAND” software. When the same seed is selected, the software produces the same result every time. In large models, this work is distributed among computing units and data streams. The researchers reconstructed the process using the same chips and software. They selected 2048 streams for each computing unit.    The resulting weight distribution exactly matched the figures in the Chinese models. When the number of computing units and streams was changed, the match disappeared completely. This allowed the researchers to determine whether “Blackwell” or “Hopper” chips were used to develop each model. They also estimated the number of cards. However, VRAM—video memory capacity—and memory speed were not taken into account. Therefore, the exact chip models are unknown.    The method works only with models whose weights have been released publicly. The result also depends on the software selected during training. In theory, “AMD” chips could emulate groups of 148 or 152 computing units. However, when the actual weight values of the “DeepSeek V4” model were tested, “AMD” hardware did not produce the same result. There is currently no explicit ban on Chinese companies renting “Blackwell” hardware abroad.