At the end of 2023, with almost no warning, Nvidia launched a "brand-new" flagship gaming graphics card on its official website in China - RTX 4090D (hereinafter referred to as 4090D). In Nvidia's explanation, this "D" version is to celebrate the Year of the Dragon. But if you are more familiar with the graphics card circle, you should know the main reasons for the birth of this graphics card:
In 2023, the U.S. Department of Commerce issued a new round of technology embargo, prohibiting the sale of some GPUs that can be used for AI and HPC to China, and the full-blooded version of the RTX 4090 is on the restricted list. As soon as this news came out, it set off an uproar in the domestic installation circle, and the AIC brands quickly "withdrew" their inventory products that had not been sold on the e-commerce platform, and even the price of 4090 on the second-hand market also rose, even exceeding 20,000 at one point.
Image source: Nvidia
Although the 4090 was actually removed from the restricted list in subsequent SEC filings, considering the uncertainty of the future, some AIC companies still decided not to sell RTX 4090 graphics cards in the Chinese market "just to be on the safe side". On the one hand, there is an uncertain red line of sanctions, and on the other hand, there is a huge game and HPC market, and under the trade-off, there is also a 4090D Chinese special version of "self-deprecating martial arts".
1. What are the similarities and differences between the two hardware?
In terms of hardware configuration, the cores of the 4090 and 4090D are AD102-300 and AD102-250, respectively.
In terms of the number of CUDA units, the 4090D is about 11% less than the 4090, from the original 16384 to 14592, and the base frequency has been slightly improved, from 2235MHz to 2280MHz, but the acceleration frequency of both is also 2520MHz. It's a bold guess that the "little overclocking" is probably to balance the graphics performance penalty that comes with CUDA reduction.
Image source: Nvidia
In terms of power consumption, both the 4090 and 4090D have a standby power consumption of 19W, but the 4090D's average gaming power consumption is slightly lower. It was reduced from 315W to 302W and TGP was also lowered from 450W to 425W.
But outside of NVIDIA's table, we also found some new changes: the full-blooded version of the 4090 has 512 Tensor Cores, but the 4090D only has 456, and the 4090 has a TPP of 5285 TOPS, while the 4090D TPP is 47.6.8 TOPS.
2. Does it have a big impact on gamers?
It has a certain impact on high-quality players.
In terms of game performance, the impact of the number of CUDA cores on the game is mainly reflected in the graphics rendering and physics simulation. The quality and complexity of graphics in games tend to require a lot of computational resources. CUDA cores are specifically designed to handle these computational tasks, including rendering high-quality graphics, real-time physics, ambient lighting and shadow calculations, and more.
When the number of CUDA cores decreases, the GPU's ability to handle these tasks decreases accordingly. This means that under all things being equal, a GPU with fewer CUDA cores may not be able to run graphics-intensive games at high resolutions or high frame rates. In addition, physics simulations are also affected. Computationally intensive physics such as explosion effects, dynamic destruction, and fluid dynamics in the game may not be accurately represented with fewer CUDA cores.
Image source: Nvidia
In addition, the changes in Tensor Core may also affect the performance of the 4090D in DLSS and frame generation, as this uses NVIDIA's AI computing.
3. Is there a way to open the core?
No.
Since the release of 4090D, there has been no stop to the discussion about cracking 4090D and "opening the core" for 4090D. "Open core" refers to the activation of additional cores in a multi-core processor, such as a CPU or GPU, to increase its processing power and performance. This concept is generally applied to processors that do not have all cores enabled from the factory. For example, some processors may have some cores disabled at the factory due to a defect in the manufacturing process or to meet the needs of different market positioning.
Image source: Nvidia
Open-core operations involve modifying the processor's settings or firmware to unlock and enable these otherwise disabled cores. This can theoretically improve the performance of the processor, especially when doing multi-threaded processing or tasks that require a lot of parallel computing.
But the problem is that we don't know anything about the AD102-250 specifically, and the chips tend to be "destroyed" by laser fusing instead of shielding the core with software, so I don't expect anything from the 4090D's "open core".
4. Will the 4090 I have already purchased be limited by the software?
No.
The 4090 and 4090D use different hardware codecs, and the drivers of the two are not mixed. So even if you download the 4090D driver for your 4090, it will be terminated with an error during installation, but the 4090D naturally cannot install the 4090 driver, and the two have different numbers in the NVIDIA driver page.
Image source: Nvidia
By the way, even though the 4090 has been discontinued and the product page has been canceled, we can still find drivers for the 4090 full blood version on the Nvidia website. From this point of view, users who have already purchased the 4090 do not need to worry about their high-performance 4090 will stop updating in the future.
Besides, even if Nvidia does stop distributing the latest drivers on Chinese websites, we can still visit the web pages of other regions and download their drivers.
5. Can domestic graphics cards be replaced?
Not at the moment.
Although domestic graphics cards have made significant progress in recent years, they are still difficult to replace Nvidia in the field of high-performance games. The reasons for this situation are manifold.
The first is the gap between technology accumulation and innovation. As an industry giant, NVIDIA has decades of technology accumulation and R&D experience. It has a strong foundation in graphics processor design, graphics rendering technology, and working with game developers. In contrast, domestic graphics cards are still in their infancy in these aspects, and there is a significant gap with NVIDIA in terms of core technologies such as chip design, software support and driver optimization.
The second is the establishment of ecosystems. Nvidia is more than just a hardware manufacturer, it has built a complete ecosystem that includes drivers, development tools, game optimization, and more. These elements work together to ensure the performance of its graphics card in high-performance games. Domestic graphics card manufacturers have not yet formed a strong enough ecosystem support in this regard.
In addition, brand influence and market recognition are also factors that cannot be ignored. With its long-term presence in the market and excellent product performance, NVIDIA has earned widespread consumer trust and brand loyalty. Domestic graphics card brands have relatively little influence on a global scale, and it takes time to build consumer trust and recognition.
In addition, game compatibility and optimization are also key points. Most high-performance games are optimized for the mainstream graphics cards in the market, while NVIDIA's graphics cards usually get more attention and optimization due to their high market share. This makes the game run more smoothly and have a better experience when using an NVIDIA graphics card. Domestic graphics cards need to cooperate with more game developers to ensure that the game also performs well on their graphics cards.
Image source: Nvidia
However, from another point of view, the significance of domestic GPUs is not in the field of games at all, and the importance of GPUs is increasing day by day in many industrial and scientific research scenarios. Therefore, compared with the adaptation of games, the primary task of domestic GPUs is to adapt to various mainstream development, simulation and other platforms, so that they can gradually replace imported GPUs in commercial and scientific research fields, and further improve the domestic chip self-sufficiency rate is the key.
For domestic GPU companies, the first task is to meet the needs of commercial, scientific research and other fields, followed by personal needs, and it is initially estimated that they will focus on the low-end market, first meet the basic entertainment and use needs, and then move to the higher-end market after stabilizing the basic market.
In fact, in addition to Innosilicon, there are many companies such as Jingjiawei that are struggling to explore and move forward on the road of domestic GPUs, and I believe that in the near future, we will see the emergence of more domestic GPU products.
CES (Las Vegas International Consumer Electronics Show) 2024 is about to be grandly opened, and Lei Technology will send a reporting team led by "One Tianming" to Las Vegas, Nevada, USA CES 2024 on-site reporting, so stay tuned.