NVIDIA RTX 5090D v2 Cuts Memory and Bandwidth 25%

On August 12, NVIDIA officially unveiled the GeForce RTX 5090D v2 for the Chinese mainland market, replacing the earlier RTX 5090D released in January. This updated version trims both memory capacity and bandwidth by about 25% to comply with the latest U.S. export control regulations. The card carries a suggested retail price of RMB 16,499 (approximately USD 2,295).

The original RTX 5090D was introduced alongside the RTX 50 series in January as a China-specific flagship model that met U.S. export rules. While its AI computing power was reduced by roughly 29% compared to the standard RTX 5090, gaming performance remained almost intact. That version featured 21,760 CUDA cores, base/boost clocks of 2.01/2.41 GHz, 32GB of GDDR7 memory on a 512-bit bus, fifth-generation Tensor and RT cores, and an AI throughput of 2,375 TOPS versus the RTX 5090's 3,352 TOPS.

However, in April, tighter export restrictions from the U.S. government further limited NVIDIA's AI-capable chips to China, affecting both the H20 and RTX 5090D. The new RTX 5090D v2 retains the same GB202-240 GPU as its predecessor but now ships with a 384-bit 24GB GDDR7 configuration, cutting bandwidth from 1,792 GB/s to 1,344 GB/s—just under the 1.4 TB/s export threshold. The card still supports all the latest RTX 5090-series technologies and maintains a 575W TDP.

This reduction stems from disabling four of the card's 16 memory channels, leaving only 12 active, which limits it to 24GB instead of 32GB of GDDR7. Despite the changes, demand for the RTX 5090D v2 is expected to be high, and early market prices may exceed the official MSRP.