This video presents a detailed technical teardown and analysis of the Huawei Atlas 300i, an early domestic Chinese AI GPU. The card, featuring a dual-GPU design with a total of 96GB of LPDDR4X memory, was acquired for $1400 for evaluation. The teardown reveals a design philosophy focused on functionality within a specific ecosystem, rather than global competition. Key findings include the use of graphite thermal pads and Integrated Heat Spreaders (IHS) on the GPUs, a practice uncommon in modern high-performance cards which favor direct-die cooling. The cooling solution is a simple passive aluminum fin stack with two flat heat pipes, designed for server environments. The PCB is notably less dense than contemporary Nvidia cards, utilizing laptop-style LPDDR4X memory modules to achieve its high VRAM capacity. Significant compatibility issues prevent the card from working in standard motherboards, necessitating the use of a specific Huawei server for future performance benchmarking.
Deconstructing the Huawei Atlas 300i AI GPU
0:00:01 Introduction to a Domestic Chinese GPU: The video introduces the Huawei Atlas 300i, a dual-GPU AI card with a combined 96GB of LPDDR4X memory. It was purchased for $1,400 as an example of China's effort to develop high-end domestic silicon in response to US export restrictions.
0:02:29 Major Compatibility Hurdles: The card is not plug-and-play and is incompatible with standard motherboards. It is designed to work only with a specific set of Huawei servers, posing a significant challenge for widespread adoption. To conduct proper testing, a compatible Huawei server has been purchased.
0:03:26 Underwhelming Memory Performance: The card's advertised memory bandwidth is only 408 GB/s across both GPUs (204 GB/s per GPU). This is significantly lower than high-end competitors like the Nvidia RTX 6000 Pro, which boasts nearly 1.8 TB/s.
0:05:40 Server-Focused Design: The Atlas 300i is passively cooled, relying entirely on the high-pressure, forced-air environment of a server chassis. It also uses a non-standard, miniaturized 8-pin power connector, requiring an adapter.
0:09:52 Unique GPU Packaging: The teardown reveals the GPUs use graphite thermal pads and an Integrated Heat Spreader (IHS). This is a throwback design, as modern high-performance GPUs use more thermally efficient direct-die cooling.
0:12:45 Non-Standard VRAM: To achieve its 96GB capacity, the card uses 24 individual 4GB LPDDR4X memory modules, a type of memory typically found in laptops and handheld devices rather than high-performance discrete GPUs.
0:15:50 A Clear Engineering Gap: When compared to an Nvidia RTX 6000 Blackwell PCB, the Huawei card's board is far less dense and complex. This visual comparison highlights the significant gap that currently exists in engineering sophistication between Huawei and established market leaders.
0:20:36 A Simple, Cost-Effective Cooler: The cooling solution consists of a basic aluminum fin stack, two thin, flat heat pipes, and an aluminum base. This simple construction is deemed adequate for its 150W power rating within a server but is not comparable to more advanced coolers on other enterprise cards.
0:25:09 Next Steps for Testing: Due to the hardware incompatibilities, the team is waiting for the arrival of a Huawei Atlas 800 server to conduct proper performance benchmarks and evaluate the GPU in its intended environment.