Nvidia's RTX-5000 partly sold with fewer ROPs than specified

Some of the first Blackwell GPUs have gone on sale with fewer raster output stages (ROP) than specified in the technical data of the graphics chips. As Nvidia has already confirmed, the RTX 5090, the China version 5090D and the 5070 Ti are affected. There are no corresponding reports for the 5080. Concrete figures and comparative measurements are mainly available for the top model 5090. According to Nvidia documents, 176 ROPs are said to be present there, but only 168 of them are active on some cards.

This fact was discovered by a user of the Techpowerup forum, who read out the data with the generally reliable GPU-Z tool. This tool shows the correct number of ROPs for other 5090 cards in the same version. In the 3DMark Timespy Extreme benchmark, the graphics card with 168 ROPs was around 14 percent slower than other models with 176 ROPs. However, different cards were compared, each with different cooling solutions. In a statement that can be found at Hardwareluxx, among others, Nvidia only states performance losses of 4 percent. AI and other compute applications should not be affected.

As the chip manufacturer goes on to write, the problem is said to be present in "less than 0.5 percent" of the volume of GPUs already delivered. According to Nvidia, this is due to a "manufacturing anomaly", whatever that means. This is said to have already been rectified. It is likely that, as is often the case with large chips, some of the ROPs did not work and were therefore deactivated. Such functional units, which are often switched off at chip level, cannot usually be switched back on by software such as a BIOS update for the GPU.

Nvidia defers problem to card manufacturer

Nvidia recommends that affected customers contact the manufacturer of the graphics card. It is unlikely that they will be able to replace the cards at the moment, as all models in the three series mentioned are constantly sold out and are only available from private sellers at inflated prices, for example on eBay. Nvidia itself is also unlikely to be able to exchange its "Founders Edition", which it sells itself.

Several things are remarkable about this process. Firstly, Nvidia does not mention the number of ROPs in its own public technical data, but it does in the "Reviewers Guide" for hardware testers. Secondly, the card manufacturers were apparently not informed in advance about different versions of the GPUs – or they did not notice the missing ROPs. Furthermore, Nvidia must have been aware of the problem. This is indicated by the quick response to the media with concrete data such as the "0.5 percent".

The more ROPs, the faster

The Raster Operations Pipeline, formerly known as the Raster Operations Processor (ROP), plays a crucial role in image generation for all graphics applications that run on a GPU. Among other things, the calculated pixels are assembled in these units, checked for visibility in the scene and finally copied to the frame buffer. Some post-processing effects and multi-sample anti-aliasing (MSAA) are also calculated in the ROPs. The more of these units a GPU has, the faster it is at these tasks.

Nvidia's specification of a performance loss of 4 percent should refer to a purely mathematical theory of 176 versus 168 ROPs; the differences can manifest themselves differently in the combination of different tasks in the ROPs. The example of 3DMark Timespy indicates this. Apart from the eight ROPs, other parts of the GPUs are unlikely to have been switched off; the GPU-Z data from users still shows 21,760 shader units even with 168 ROPs.

(nie)