Nvidia Unveils Specialized Chips for Chinese Market Amidst US Restrictions
In a strategic move to comply with stringent US chip export regulations, Nvidia has discreetly launched two consumer-grade chips, the RTX 5880 Ada workstation GPU and the RTX 4090 D flagship graphics card. These “China Special Edition” chips come as a response to the chip giant’s efforts to navigate the restrictions imposed by the US government.
Performance Adjustments to Navigate Export Control Measures
The RTX 5880 Ada, featured on Nvidia’s official website, showcases a substantial 25% decrease in performance compared to its flagship counterpart, the RTX 6000. Equipped with a modified AD102 GPU and 14,080 CUDA cores, the RTX 5880 Ada’s single-precision floating-point performance takes a hit of approximately 24%. Despite the diminished capabilities, it positions itself as a viable alternative, approaching the performance of the RTX 5000.
Nvidia Struggle with US Export Control Measures
This move follows Nvidia’s earlier release of the RTX 4090 D, a consumer-grade flagship card, tailored for the Chinese market. The RTX 4090 D faced a 10% reduction in AI performance while maintaining a starting price of 12,999 yuan. Nvidia’s CEO, Huang Renxun, emphasized the company’s commitment to adhering to trade regulations and expressed their dedication to providing China with products compliant with the latest US government specifications.
Catering to the Chinese Market Needs
Despite the absence of explicit confirmation regarding the RTX 5880 Ada’s exclusive focus on the Chinese market, the utilization of the same performance reduction strategy as the 4090 D suggests its role in bypassing the semiconductor export control measures imposed by the US Department of Commerce.
Nvidia aims to seek market feedback as part of its ongoing process to comply with regulations. This move aligns with Huang Renxun’s confirmation that Nvidia will continue offering compliant chip products for the Chinese market, including “special edition” chips like the HGX H20 and L20 PCIe, as mentioned in the warnings from US Secretary of Commerce Gina Raimondo.
Specifications and Expected Pricing
The RTX 5880 Ada boasts 14,080 CUDA cores, 440 Tensor cores, and a clock frequency of approximately 2.5 GHz. With 48 GB GDDR6 memory running at 20 Gbps, it features a standard dual-slot active cooling design and four DisplayPort 1.4a output ports. While Nvidia has yet to disclose the pricing, it is anticipated to align closely with the RTX 6000, likely around $6,800 (approximately 48,300 RMB).
AI Training and Inference Capabilities
Industry experts suggest that while the RTX 5880 and 4090 D excel in single-precision computing power, limitations in memory and bandwidth may hinder their ability to train complex models like GPT with 700TB/trillion parameters. However, for smaller models like Llama 2-7B and 13B, a single RTX 4090D can handle stable operation, and combining eight can even train models with 70 billion to 650 billion parameters. In terms of inference, both cards perform robustly, particularly in AI graphics rendering and video generation.
Future Prospects and Continued Market Presence
Reports indicate that Nvidia plans to resume shipments of “special edition” AI chips to China, with mass production expected to commence in the second quarter of 2024, including the H20 and other AI compute chips for data center applications. Raimondo’s statements suggest that the US is open to limited sales of AI chips to China, but restrictions on the most advanced AI chips persist. Nvidia’s innovative strategies and adaptability in the face of geopolitical challenges position them as a key player in the evolving landscape of AI chip manufacturing.
