Microsoft's DirectStorage API has been known to have weird performance quirks with GPU decompression. However, YouTuber Compusemble claims that Nvidia's latest GeForce RTX 5090 GPU handles GPU decompression better than the prior generation RTX 4090, showing less of a performance gap when using GPU decompression vs CPU decompression.
The YouTuber benchmarked two games with DirectStorage (DS) support, Ratchet and Clank: Rift Apart and Spider-Man 2. The games were tested with DirectStorage GPU decompression and CPU decompression, on an RTX 5090 at 4k, 1440p, and 1080p resolutions. Note that no RTX 4090 tests were shown for comparison, which has an a priori assumption that DS causes a loss of performance.
In Ratchet and Clank: Rift Apart, the RTX 5090 produced nearly identical average frame rates with GPU decompression compared to CPU decompression at native 4k. Average frame rates were just 0.96% faster with CPU decompression, while 1% lows were almost 10% higher in favor of CPU decompression. That suggests that, once again, a higher GPU load (e.g. at 4K) will cause DS running on the GPU to potentially reduce performance.
1440p flips the story, and while it showed slimmer margins, this time overall performance (specifically the 1% lows) favored GPU decompression. DirectStorage GPU decompression on the RTX 5090 provided 1.19% higher frame rates on average, and 5.26% better 1% lows than CPU decompression. 1080p saw GPU decompression winning again, featuring 0.48% higher frame rates on average and 11.11% better 1% lows.
On the other hand, Spider-Man 2 demonstrated a performance bias against GPU decompression at all three resolutions. At 4k, the RTX 5090 performed 8.2% better in average FPS and 10.53% better in the 1% lows with CPU decompression. At 1440p, CPU decompression performed 3.53% better in average FPS and 14.29% better in 1% lows. Finally, at 1080p, CPU decompression yielded 3.3% higher average FPS and 9.23% higher 1% lows compared to GPU decompression.
Consumable claims the RTX 5090's performance is more consistent with GPU and CPU decompression compared to the older RTX 4090. An older video he recorded backs up his statement, showing the RTX 4090 with a greater performance drop-off when toggling between CPU and GPU decompression, specifically in Spider-Man 2. With the RTX 4090 at 4k in Spider-Man 2, average frame rates were 10.34% higher and 1% lows 17.95% higher with CPU decompression. At 1440p, the average frame rate was 6.25% higher and 1% lows 18.87% higher with CPU decompression. At 1080p, the average frame rate was 3.25% higher and 1% lows 25.86% higher with CPU decompression.
Considering the RTX 5090 has both more raw compute (105 TFLOPS FP32 vs 83 TFLOPS) as well as significantly more memory bandwidth (1.8 TB/s vs 1.0 TB/s), as well as 33% more VRAM, those could all contribute to the reduce impact of GPU decompression. The RTX 5090 boasts a 512-bit wide memory interface and 28 Gbps GDDR7 memory modules, giving it 78% more memory bandwidth than the RTX 4090. GPU decompression is very heavy on the memory system, since assets need to be streamed into GPU memory for the decompression process to function.
Of course, shader computations still come into play, and so it makes sense that, at lower resolutions where the 5090 is more likely to be completely CPU limited, it would also have the spare cycles to handle decompression. But whatever the cause (architecture could also play a role), the 5090 doesn't seem to mind GPU decompression as much as the 4090. Still, what we want to see is more games using DirectStorage to improve load times as well as overall performance — slightly longer load times would otherwise be preferable to inconsistent framerates.
