Conclusion
We have now pitted AMD’s flagship GPU to NVIDIA’s flagship GPU in compute, content creation, professional workstation application benchmarks. We have also recently fully reviewed the AMD Radeon RX 6900 XT in many games at 1440p, 4K, with Ray Tracing and FSR. We also overclocked it. These compute benchmarks today should provide the full scope, combined with the gaming performance as to how the Radeon RX 6900 XT and GeForce RTX 3090 compete. Depending on your work case and workload scenario, you should be able to decide which video card will provide the best experience for you.
Compute Performance
We started off in Geekbench 5 looking at compute OpenCL and Vulkan performance. Both OpenCL and Vulkan compute tests were much faster on the RTX 3090. OpenCL by 23% and Vulkan by a much larger 57%. We were also able to test CUDA, and that further improved performance over OpenCL for the RTX 3090. Next, we moved on to LuxMark and tested Path Tracing with Global Illumination and with just brute force Path Tracing. In both scenarios, the RTX 3090 came out on top by 20-25%. We then tested video transcoding with HandBrake, and while it did help processing time over the CPU both video cards actually performed it in a similar time. There was no advantage toward either one.
We ran the full suite of AIDA64 GPGPU benchmarks to test functionality between both GPUs. It started off showing us that memory read, write and copy was much faster on the Radeon RX 6900 XT versus the RTX 3090. Then we had performance jump back and forth. It seems Single-Precision FLOPS was 45% faster on the RTX 3090, but Double-Precision FLOPS were 147% faster on the RX 6900 XT. We also found 24-bit Integer to be faster on the RX 6900 XT. However, once we moved up in precision to 32-bit and 64-bit Integer the RTX 3090 stole the show. AES-256 and SHA-1 Hash were also 30-50% faster on the RTX 3090. Single-Precision Julia was 30% faster on the RTX 3090, but Double-Precision Mandle was 209% faster on the RX 6900 XT.
In Blender Benchmark the consistent result was that the RTX 3090 with CUDA rendered the scenes generally faster than the Radeon RX 6900 XT in OpenCL. There was only one instance where the RX 6900 XT was faster than the RTX 3090 with CUDA, but it was not a major win. What was the major win was turning on and using OptiX on the RTX 3090, that one stole the show on rendering performance in Blender. It’s night and day with the RX 6900 XT in OpenCL.
In the PerformanceTest 10 PassMark Advanced GPU Compute tests we tested various DirectCompute and OpenCL tests. In the NBodyGravity physics test, the RTX 3090 came out on top by 25%. However, in Fluid Simulation the RX 6900 XT was faster slightly by a few percent. In the Julia Fractals, the RTX 3090 was faster by 51%, but in Madelbrot it was faster by only a few percent. The OpenCL physics test here shows the RX 6900 XT being 51% faster than the RTX 3090.
Moving on to SPECviewperf 2020 seems to mostly favor the RTX 3090 in performance. It is anywhere up to 52% faster depending on the application. There are, however, a couple of applications that really prefer the Radeon RX 6900 XT. In energy-03 the RX 6900 XT was 18% faster and in medical-03 the RX 6900 XT was 39% faster and in snx-04 it was 260% faster. Surely if you are using those apps, go with the RX 6900 XT.
Final Points
Generally speaking, it seems that in lower Integer precision the Radeon RX 6900 XT is faster. However, as integer precision increases, the RTX 3090 takes huge leaps of performance over the RX 6900 XT. It also seems single-precision is faster with RTX 3090, while RX 6900 XT seems to handle double-precision better. Moving data in and out of VRAM seems to be quicker on the RX 6900 XT. Things like encryption, however, seem to favor the RTX 3090.
For Blender rendering, you really cannot beat CUDA on the RTX 3090 and especially OptiX. If you Blender, that’s the way to go for sure. If you are doing any kind of rendering with Path Tracing, the RTX 3090 will be faster. Video transcoding actually seemed to be very similar between the two. This one will specifically depend on the application you are using and if it benefits from VCE, or CUDA, or NVENC better.
In terms of professional workstation applications, it seems the RTX 3090 is mostly the best GPU for the job. Unless of course, you are using those three specific applications that were just flat out better with the Radeon RX 6900 XT in a big way.
Overall which GPU is better suited for your workload will depend heavily on what that workload is. There are scenarios where the Radeon RX 6900 XT will be better, and scenarios where the RTX 3090 is better suited. Generally speaking, it seems from our testing that the RTX 3090 is more well-rounded for compute, content creation, and professional workstation applications, while the Radeon RX 6900 XT is more specifically tuned to specific workloads.