China's GPU Competition: 96GB Huawei Atlas 300I Duo Dual-GPU Tear-Down

Summary

This video dives into the detailed examination of a high-performance GPU, designed primarily for servers and heavy workloads. The review highlights the GPU’s cooling system, which uses thin heat pipes and aluminum fins, and its hand-assembled construction, which prioritizes manual assembly for heat sink installation. The GPU features 96GB of VRAM, with a focus on memory cooling that’s less than ideal but sufficient for the task. Despite being priced around $1,400, the card’s cooling system is noted to be minimal compared to other high-end GPUs. The review also discusses upcoming testing in server environments, with a focus on its compatibility and performance under real-world conditions.

Takeaways

• 😀 The GPU cooling system relies on a combination of air pressure and airflow to ensure efficient heat dissipation, with a focus on the fin stack and airflow dynamics.
• 😀 Thin heat pipes used in the GPU design are unlikely to contribute significantly to cooling, especially in comparison to other more robust cooling solutions.
• 😀 The card’s memory setup consists of 6 modules per GPU, which contributes to a lower memory density than other high-end GPUs.
• 😀 The card has 96 GB of VRAM in total, with 4 GB modules per memory channel, although the bandwidth is not as high as more modern configurations.
• 😀 The backplate of the GPU features thermal pads that only partially cover the memory, which is not ideal but likely sufficient for the card's power demands.
• 😀 Hand assembly of the cooler components (such as heat sinks) is typical in the video card manufacturing process, especially for parts like thermal solutions and some inductors.
• 😀 The cooling design on this card uses an aluminum fin stack rather than copper, which suggests cost-saving measures in the cooling solution.
• 😀 Despite its high $1,400 retail price, the card's cooling solution does not meet the performance level of some higher-end models, suggesting trade-offs in cost vs. performance.
• 😀 There is some limited testing being planned, with the card to be tested in a server environment, though compatibility with specific systems is still uncertain.
• 😀 The host plans to test GPUs from Chinese manufacturers, providing insight into the emerging global competition in GPU silicon, though official support for these cards in certain servers is limited.

Q & A

What is the purpose of the aluminum fin stack in the GPU cooling system?

The aluminum fin stack helps to dissipate heat from the GPU by forcing air through the fins under pressure. This ensures efficient heat removal, even though the fins are not particularly dense. The system is designed to pull heat from the GPU and pass it through the fins to keep it cool.

Why are the heat pipes described as 'ultra-thin' and what effect does this have?

The heat pipes in the cooling system are described as ultra-thin, similar to those used in laptop-style coolers. These thin heat pipes may not provide as much cooling efficiency as thicker ones, but they are likely chosen due to space constraints and weight considerations.

What does the 'witness mark' on the base and screw indicate?

The 'witness mark' on the base and screw indicates that the assembly process is likely hand-assembled. This mark helps the assembly team know how much to tension the screws, ensuring proper installation of the cooling components.

How many memory modules does the GPU have, and what is the total VRAM?

The GPU has 24 memory modules, with 6 modules per GPU. Each module is 4GB, bringing the total VRAM to 96GB.

What type of memory is used in the Blackwell card compared to this one?

The Blackwell card uses a more modern GDDR7 memory, which is likely to have higher density and better bandwidth compared to the older GDDR5 or GDDR6 memory used in this GPU. The Blackwell card uses fewer total modules but with higher density and bandwidth.

Why is the memory cooling solution considered less optimal?

The memory cooling solution is considered less optimal because only the centers of the memory modules are covered with thermal pads, rather than covering the entire pad. While this is not ideal, it may still be sufficient for the GPU's cooling needs, especially considering that the memory modules are likely low-power.

What does the simplicity of the GPU’s cooling solution indicate about its design?

The simplicity of the GPU's cooling solution, which consists of just two heat pipes and aluminum sinks, suggests that the cooling may not be as robust or high-end as other more expensive cards. The design likely focuses on efficiency rather than premium performance.

What does the reference to hand-assembly in the cooling process imply?

The reference to hand-assembly in the cooling process implies that many components, such as heat sinks and certain screws, are installed manually rather than by automated machines. This is common in the assembly of GPUs, especially for the more intricate cooling solutions.

What is the significance of testing this GPU in a server with a Threadripper CPU?

Testing the GPU in a server with a Threadripper CPU is important because it may not be officially supported by the server, which could influence the GPU's performance. The testing will provide insights into how well the GPU works in an unconventional setup, especially with different server CPUs like the Chinese-made Quinn Pong CPUs.

What are the potential challenges of running this GPU in the Chinese-made servers?

The main challenge of running this GPU in Chinese-made servers is the limited Nvidia and AMD support. These servers have a short list of officially supported GPUs, which could result in compatibility issues. Additionally, the GPU's performance might be impacted by the server's hardware limitations.