Nvidia’s RTX 5090D V2 Downgraded for China with 25% Less VRAM and Bandwidth, Same High Price

Nvidia has quietly launched the GeForce RTX 5090D V2, a China-exclusive variant of its flagship Blackwell-based graphics card, designed to comply with U.S. export restrictions. The new model represents a further downgrade from the original RTX 5090D, featuring a 25% reduction in VRAM and memory bandwidth, despite maintaining the same core architecture and price point. The RTX 5090D V2 comes with 24GB of GDDR7 memory—down from 32GB on the standard RTX 5090 and RTX 5090D—along with a reduced 384-bit memory bus, down from 512 bits. This change results in a 25% drop in memory bandwidth, from 1,792 GB/s to 1,344 GB/s. The reduction stems from four disabled memory channels, meaning the card only connects to 12 of the 16 available memory chips on the GB202 die, leaving four empty pads on the PCB. Despite these cuts, the GPU’s core specifications remain largely unchanged. The card retains the same 92.2 billion transistors, 750 mm² die size, 170 streaming multiprocessors, 21,760 CUDA cores, 680 Tensor Cores, and 170 Ray Tracing Cores as its predecessors. It also maintains the same 2,407 MHz boost clock and 104.8 TFLOPS of FP32 performance. However, AI performance has been capped at 2,375 AI TOPS—down from 3,352 on the full RTX 5090—aligning with U.S. export controls that limit AI capabilities in chips sold to China. The L2 cache remains at 96MB, consistent with the original models, though Nvidia has not confirmed whether this remains unchanged on the V2 variant. Given that the full GB202 die includes 128MB of L2 cache, and previous cards already used only a portion of the silicon, the cache size likely hasn’t been altered. While the performance impact on gaming may be noticeable, especially in memory-intensive titles, the real consequence lies in AI workloads. With reduced VRAM and bandwidth, the RTX 5090D V2 is significantly less effective for training and running large AI models—precisely the use case driving demand in China’s AI market. What’s particularly frustrating for consumers is that the RTX 5090D V2 is priced at $2,299, the same as the original RTX 5090D, despite offering substantially less performance and capacity. In today’s volatile market, where retail prices often exceed MSRP, buyers may end up paying even more for a less capable product. The move underscores the growing challenge of balancing global market access with geopolitical restrictions. While Nvidia continues to adapt its products for regulated regions, Chinese users are left with a high-cost, lower-performance alternative—effectively a downgraded flagship.