Huawei AI GPU has the same memory size as RTX 6000 Blackwell, but costs nearly 6 times less
China blocked from high-performance NVIDIA GPUs is already making their own alternatives.
Gamers Nexus has taken Huawei’s Atlas 300I Duo apart, a dual-GPU accelerator with a total of 96 GB of LPDDR4X memory (48 GB per chip). For them it was priced around $1,400 after shipping, but only because they used unofficial channels. This card is not available for sale outside China, the same way NVIDIA is blocked from selling its GB202-based RTX 5090 non-D and RTX PRO 6000 cards to that country.
Source: Gamers Nexus
The Huawei card costs a fraction of NVIDIA’s RTX 6000 Blackwell Pro, which sells for about $8,000, but the two are not in the same performance class. Huawei’s solution targets domestic servers and currently faces hardware and software compatibility issues, working only with select Huawei platforms.
Source: Gamers Nexus
The Atlas 300I Duo offers just 204 GB/s of bandwidth per GPU, compared to 1.8 TB/s on NVIDIA’s workstation flagship. It uses simple aluminum cooling, and an LPDDR-based design rather than modern GDDR or HBM memory. Still, it represents a major domestic effort to build a Chinese AI GPU ecosystem amid export restrictions on high-end NVIDIA models such as the RTX 6000 Pro Blackwell. It’s main feature is not the speed or form factor, but the memory capacity, very much needed for AI training.
Source: Gamers Nexus
The teardown shows a simple design compared to NVIDIA’s packed workstation boards. Huawei’s layout, power delivery, and cooling setup look basic, using graphite pads, thin heatpipes, and an aluminum heatsink. The card runs at around 150W and doesn’t have its own fans, so it needs strong airflow from a server to stay cool.
To test the Atlas 300I Duo, Gamers Nexus is now sourcing Huawei’s Atlas 800 server equipped with domestic Kunpeng 920 CPUs, as the GPU cannot run in standard desktop systems. I guess they will have another video, and it looks like Steve will also do a teardown of the IHS later.
Source: