Huawei's Atlas 300I Duo packs a 96GB LPDDR4X AI GPU at $1,400 for servers. Compare it to NVIDIA RTX 6000 Blackwell and see why memory capacity trump speed.
© Gamers Nexus
Huawei has unveiled its own AI GPU, the Atlas 300I Duo, which many are already pitching as an alternative to NVIDIA’s pricey cards. The board carries 96 GB of LPDDR4X memory, split across two chips at 48 GB each, and is priced at about $1,400—nearly six times less than the $8,000 NVIDIA RTX 6000 Blackwell Pro. Despite the identical memory footprint, the two products aren’t direct rivals, and the comparison feels more about budgets than like-for-like performance.
Gamers Nexus tore the card down and found a straightforward design inside: an aluminum heatsink, slim heatpipes and graphite thermal pads. The board draws around 150 W and has no onboard fans, relying on server airflow for cooling. While NVIDIA leans on high-speed GDDR or HBM, Huawei goes with LPDDR4X. The trade-off is clear: a memory bandwidth of 204 GB/s versus 1.8 TB/s on the RTX 6000 Blackwell, which helps keep costs down but narrows throughput. In other words, Huawei prioritizes capacity and affordability over raw speed.
Atlas 300I Duo is positioned as a server accelerator and isn’t compatible with standard desktop PCs. For testing, it’s paired with Huawei’s Atlas 800 server running Kunpeng 920 processors. Here the card’s main advantage stands out—the sheer memory capacity, a critical factor for training neural networks. For workloads that hinge on fitting larger models or datasets, that can be more decisive than headline bandwidth.
Experts note that the Atlas 300I Duo trails NVIDIA’s offerings in speed and energy efficiency, yet its price and availability within China make it a cornerstone in building an independent ecosystem of AI accelerators. Against the backdrop of export controls on NVIDIA shipments to China, Huawei’s approach looks less like a head-to-head duel and more like a pragmatic pillar for the local industry.