Improved RTX 5090D V2 performance shows that a 384-bit memory could suffice, with the Chinese version performing nearly identical to the original in specific benchmark tests.
In a surprising turn of events, Nvidia's latest graphics card, the RTX 5090D V2, has shown near-identical gaming performance to its predecessor, the RTX 5090D, despite having about 33% less memory bandwidth.
According to benchmarks conducted by Yesky, the RTX 5090D V2 performed on par with the RTX 5090D in various gaming titles such as Cyberpunk 2077, Hogwarts Legacy, Borderlands 3, Far Cry New Dawn, and Marvel Brawl. In Cyberpunk 2077 with DLSS 4 at 1440p, the RTX 5090D produced 424 FPS and the RTX 5090D V2 produced 418 FPS. In Far Cry: New Dawn at 4k resolution, the RTX 5090D produced 197 FPS and the RTX 5090D V2 produced 195 FPS.
The performance scores of the RTX 5090D V2 are also similar to the RTX 5090D in benchmarks such as 3DMark TimeSpy Extreme, Fire Strike, Fire Strike Ultra, Speed Way, Port Royal, and Steel Nomad. For instance, in 3DMark TimeSpy Extreme, the RTX 5090D scored 25,366 and the RTX 5090D V2 scored 25,105. In 3DMark Fire Strike, the RTX 5090D scored 108,989 and the RTX 5090D V2 scored 106,601.
So, what accounts for this unexpected consistency in gaming performance? It appears that Nvidia's Blackwell architecture does not significantly benefit from the wider 512-bit memory bus in gaming tasks. Reducing the bus width to 384-bit and lowering memory bandwidth from about 1790 GB/s to 1340 GB/s has little impact on real-world gaming performance.
This is due to several factors. First, architectural limitations prevent the GPU from fully utilizing the larger 512-bit memory interface for gaming workloads, even at 4K resolution. Second, memory efficiency improvements such as larger or more efficient caches, and faster GDDR7 memory help compensate for the narrower memory bus. Lastly, the 5090D V2 was designed to comply with export restrictions in China that limit memory bandwidth to 1.4 TB/s, explaining the memory bandwidth reduction and VRAM cut.
While gaming performance remains close, workloads that are heavily dependent on memory bandwidth, such as AI or professional compute tasks, show more noticeable performance degradation on the 5090D V2 due to the reduced memory bandwidth.
Interestingly, Yesky's benchmarking results suggest that Nvidia's Blackwell architecture cannot take advantage of the RTX 5090 and RTX 5090D's 512-bit memory interface for gaming, even at 4k resolution. This trend is not new, as the RTX 4090, RTX 3090, and RTX 2080 Ti all featured 384-bit memory interfaces in the past.
Despite the lowered memory specs, the RTX 5090D V2 costs the same as the outgoing 5090D. It comes with a memory configuration of 24GB GDDR7 and a 384-bit interface, similar to the RTX 4090 and RTX 3090. The RTX 5090D V2 is the third RTX 5090 variant and the second RTX 5090D model.
In conclusion, the RTX 5090D V2 achieves near-identical gaming performance to the RTX 5090D due to the limitations of Nvidia's Blackwell architecture and the typical gaming workloads, which do not fully exploit the maximum theoretical memory bandwidth of a 512-bit bus. This allows Nvidia to reduce memory specifications without hurting gaming benchmarks significantly.
