A few days ago, Nvidia admitted that its insane performance claim of the new $549 RTX 5070 being equivalent to the $1499 RTX 4090 was only possible with DLSS 4. That is because the company shared additional performance slides alongside the ones it shared at its CES 2025 public event.
A senior Nvidia official revealed that the RTX 5080 was going to be around 15% faster than the 4080 and that we should also expect a similar performance bump from the RTX 5090 as well versus the 4090. That was in the case of ray tracing, though, as Nvidia did not directly disclose the rasterization improvement.
Today, the Geekbench numbers for the RTX 5090 have leaked and it is impressive as well as a bit disappointing at the same time.
Keep in mind that we can't directly compare Geekbench results with gaming performance as it is a compute workload. However, it can give us a rough estimate of the performance improvement in terms of how all the extra TFLOPS and the IPC (instructions per clock/cycle) on a GPU can help it.
We start with the mediocre as we have the OpenCL data first where the RTX 5090 scored up to 367,740 points. This was the best score out of three runs and they show a lot of variance. The lowest score was 331,525. The best score is ~15.87% better than an RTX 4090 which does an average of 317,365.
Up next, we have the Vulkan benchmark and in this test, the RTX 5090 performs much better. Here the variance of the three scores was also not as equally spaced apart as it only underperformed in one out of the three runs.
The highest score the RTX 5090 managed to put up is 359,742. In comparison, the RTX 4090 on average does 262,487, which is an improvement of ~37.05%. Thus, it looks like the new Blackwell architecture brings some special improvements and optimizes related to Vulkan GPU compute that was missing in the previous generation.
We should know more about the true ability of the RTX 5000 series GPUs once the third-party benchmarks are out.
Source: Geekbench (1 | 2) via BenchLeaks (X) (1|2|3|4|5|6)
Update: The RTX 5090 was also tested in CUDA on Geekbench 5 and the GPU scored 542,157. In comparison, the RTX 4090 can do about ~415,145; thus, in CUDA, the 5090 is ~30.6% faster (via BenchLeaks).
13 Comments - Add comment