Source: Titanium Media

Image Source: Generated by Wujie AI

In order to meet the US export control requirements for chips to China, NVIDIA has made great efforts and exquisite techniques.

On January 6th, NVIDIA, the chip giant, quietly launched the NVIDIA RTX 5880 Ada workstation graphics card on its official website, mainly targeting chip products for consumer and professional fields such as AI training and inference.

Compared to the flagship RTX 6000, the NVIDIA RTX 5880 has significantly reduced performance, using a castrated AD102 GPU with 14,080 CUDA cores, a 22% reduction in the number of CUDA cores compared to the RTX 6000, and a decrease of about 24% in single-precision floating-point performance, resulting in an overall performance reduction of nearly 1/4, with actual performance approaching that of the RTX 5000, which is the next flagship.

Comparison between NVIDIA RTX 6000, 5880, and 5000 graphics cards

At the end of 2022, NVIDIA officially launched the long-rumored "China Special Edition" consumer flagship graphics card, the RTX 4090 D, with a 10% reduction in AI performance and a starting price of 12,999 yuan.

It is understood that as of now, some server agents in China have obtained samples of the 4090D and test version graphics cards.

Now, NVIDIA has officially released the RTX 5880 Ada on its official website, indicating that it has begun to accept customer purchases of this product. It is worth noting that although NVIDIA has not specifically stated that the RTX 5880 is targeted at the Chinese market and is displayed and sold in the global market, given that it uses the same "castrated" performance reduction method as the 4090D, there is reason to believe that the RTX 5880 Ada is a product aimed at circumventing the semiconductor export control measures issued by the US Department of Commerce.

"We established the company to do business and strive to do business with all possible people," said NVIDIA CEO Huang Renxun recently, stating that the company will continue to "perfectly" comply with trade regulations and provide a set of new products that comply with the latest regulations of the US government for the Chinese market. He added that NVIDIA needs to seek market advice, and this process is ongoing.

On December 6, 2023, under the public warning of US Secretary of Commerce Gina Raimondo, Huang Renxun confirmed that NVIDIA will continue to provide compliant chip products for the Chinese market. It is expected that the "special supply" chips will include products such as HGX H20, L20 PCle, and L2 PCle.

Subsequently, NVIDIA China officially released the GeForce RTX 4090 D on its official website, a customized version designed to cope with the US "chip ban," with performance lower than the export control standard set by the US.

Now, the NVIDIA RTX 5880 Ada has also been officially released. In terms of specifications, the NVIDIA RTX 5880 Ada graphics card has 14,080 CUDA cores and 440 Tensor cores, with a main frequency of about 2.5 GHz, providing 69.3 TFLOPs of FP32 computing power and 1108 TFLOPs of Tensor performance. Compared to the RTX 6000 Ada, the FP32 and Tensor core performance of the RTX 5880 is reduced by 24%; in terms of memory, the RTX 5880 uses 48 GB of GDDR6 memory with a running speed of 20 Gbps, a bandwidth of 960 GB/s, and a standard dual-slot active cooling design, with four DisplayPort 1.4a output ports.

Although NVIDIA has not disclosed the pricing information for the RTX 5880, it is expected that the price of the RTX 5880 Ada will be similar to that of the RTX 6000, with a selling price of around $6,800 (approximately 48,300 RMB).

Regarding whether the NVIDIA consumer-level RTX 5880 and 4090D graphics cards can be used for AI model training and inference, industry insiders revealed to the Titanium Media App that the large-scale AI training of graphics cards mainly depends on computing power, memory, and bandwidth capabilities. The RTX 5880 and 4090D are generally excellent in terms of single-precision computing power, but are limited in terms of memory and bandwidth, and cannot train models with parameters as large as GPT's 700TB/trillion. However, for small parameter models such as Llama 2-7B and 13B, a single 4090D card can run stably. If eight 4090D cards are combined, they can also train models with 7-65 billion parameters. In terms of inference, based on the Ada architecture and CUDA software, the RTX 5880 and 4090D can run stably, especially in AI graphics rendering and video generation, reaching a "top-level" performance.

At present, the NVIDIA RTX 5880 and 4090D graphics cards will become powerful computing chip products that very few domestic enterprises can purchase and can stably run AI model training and inference.

According to "Reference News," NVIDIA will resume the shipment of "special supply" AI chips to China, and it is expected that mass production of H20 and other AI computing chips for the data center field will begin in the second quarter of 2024. Raimondo has stated that the US can allow NVIDIA to sell AI chips to China to a limited extent, but does not allow NVIDIA to export the most complex and powerful AI chips.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。