After new US sanctions on GPU hardware effectively banned NVIDIA's flagship GeForce RTX 4090 from being sold in China, it didn't take long for the company to quietly announce a special cut-down model for the region called the GeForce RTX 4090D.
With fewer CUDA Cores, RT Cores, and Tensor Cores compared to the baseline GeForce RTX 4090 but the same price, we now know the impact on performance. And it's not too bad, and certainly not a massive nerf.
With a new GALAX RTX 4090 D Metal Master card put to the test by Expreview, it turns out that it's only 5.5% slower on average for rasterized gaming, which increases slightly to 5.8% when turning on ray-tracing. However, this gap drops to 5% when DLSS 3 Frame Generation is enabled. Either way, 5% makes the new GeForce RTX 4090 D similar to the RTX 4090 - with the difference barely noticeable for gaming workloads.
The outlet carried out benchmarks across 16 games, so it's a pretty comprehensive review. The RTX 4090D is 6% slower for AI and productivity, so it's still decent. However, in Stable Diffusion and Blender, there was around a 10% drop-off in performance in certain tests.
The GeForce RTX 4090D features 14,592 CUDA Cores compared to 16,384 in the standard RTX 4090 - a 10% reduction. This 10% reduction also applies to the GPU's RT and Tensor Cores. The RTX 4090D uses a different chip, AD102-250, compared to the 4090's AD103-300 - which also sees the TDP lower to 425W. Memory speed and capacity are the same.
One interesting tidbit is that the outlet was able to overclock the RTX 4090D by 200MHz, so the idea that NVIDIA would put a lock on overclocking these new cards to meet the demands of existing sanctions is false.