The video provides a detailed teardown of the Huawei Atlas 300i Duo, a dual-GPU AI card with 96 GB of memory that, despite its impressive capacity, falls short of Nvidia’s performance and features due to early-stage development and export restrictions. It highlights China’s efforts to build a domestic GPU industry amid geopolitical challenges, showcasing the card’s server-focused design, modest cooling, and lower memory bandwidth compared to Nvidia’s high-end GPUs.
The video presents an in-depth teardown and analysis of one of China’s earliest domestic high-end AI GPUs, the Huawei Atlas 300i Duo, a dual-GPU card with a combined 96 GB of memory split evenly between two GPUs. Although not officially available in the American market, the reviewers acquired the card through unconventional means for $1,400, which is significantly cheaper than Nvidia’s 96 GB RTX Pro 6000 Blackwell priced at around $8,000. Despite its impressive VRAM capacity, the Huawei card’s performance is currently far behind Nvidia’s offerings, reflecting the early stage of China’s GPU development amidst export restrictions and geopolitical tensions.
One of the main challenges highlighted is the card’s limited compatibility with standard consumer motherboards, as it is designed specifically for certain Huawei servers. This creates hurdles in testing and benchmarking the card outside its intended environment. To address this, the team purchased a Huawei Atlas 800 inference server equipped with domestic Chinese CPUs to properly test the GPU in a supported platform. The card lacks active cooling, relying instead on forced airflow from server chassis cooling solutions, which is typical for server-grade hardware but unusual for consumer GPUs.
Technically, the Huawei Atlas 300i uses ECC LPDDR4X memory, which is uncommon in discrete GPUs and offers lower bandwidth compared to Nvidia’s high-end cards that utilize faster GDDR7 memory. The card’s memory bandwidth per GPU is around 204 GB/s, significantly lower than Nvidia’s 1.8 TB/s. The GPU also features an integrated heat spreader (IHS) on the silicon die, a design choice more common in CPUs than modern GPUs, which typically use direct die cooling for better thermal performance. The cooling solution itself is relatively simple, with two flat heat pipes and an aluminum fin stack, reflecting a more modest engineering approach compared to Nvidia’s densely packed, high-performance boards.
The teardown also reveals that the card is hand-assembled with some thermal pads and simpler VRM components, contrasting sharply with the dense and sophisticated PCB designs found in Nvidia’s high-end GPUs. The card is rated for 150 watts and uses smaller power connectors typical of server hardware rather than the larger PCIe or EPS connectors common in consumer GPUs. Overall, the build quality and cooling solutions indicate that Huawei is still in the early phases of competing in the high-end GPU market, focusing primarily on serving the domestic Chinese market under export restrictions.
In conclusion, the Huawei Atlas 300i represents an important step in China’s efforts to develop domestic GPU technology amid geopolitical and trade challenges. While it currently lags behind Nvidia in performance, software support, and ecosystem maturity, it showcases China’s ambition to build a self-reliant AI hardware industry. The reviewers plan to conduct further testing once the appropriate server arrives, offering a rare technical insight into China’s emerging GPU market. This development is part of a broader narrative involving export controls, government funding, and the race for semiconductor independence between China and the US.