Why you can trust Tom's hardware?
Our expert reviewers spend time testing and comparing products and services so you can choose the one that's best for you. Learn more about how to test.
Games aren't the only modern GPUs. They are used to offload tasks such as video encoding from the CPU to accelerate professional CAD/CAM and scientific applications, and are particularly useful for AI. We have refined our expert and AI test suites to explore the various GPUs in greater detail. This is because AI benchmarks tend to be more important to a wider range of users.
image 1 of 3
Procyon has multiple AI tests and ran the AI vision benchmark along with two different stable diffusion image generation tests. There are several variations of this test that have been determined to be roughly equivalent (output) to UL: OpenVino (Intel), Tensort (Nvidia), and DirectML (mainly for AMD). Some tests also have options for FP32, FP16, and INT8 data types. This gives different results. We tested the available options and used the best results for each GPU.
nvidia usually flocks AMDs in Procyon AI tests, but the RTX 5060 TI 8GB has some issues, just like the RTX 5060. Neither can run a stable diffusion XL workload at the expected rate. In the past, Nvidia has said that it still appears to have a twist in the drivers working out on the Blackwell RTX 50 Series GPU, and this is a great example of that.
In AI vision tests, the fastest AMD cards still can't even match the RTX 3060, so don't worry about the RTX 5060 and 5060 Ti. Intel's ARC B580 is far better than anything else in AMD, but AI Vision doesn't work in integer mode via DirectML. Stable diffusion 1.5 represents a more typical generation AI workload, but not a lot of VRAM required. Here, the 9060 XT essentially ties the RX 7800 XT, but it's behind the ARC B580 and RTX 5060.
And it becomes stable diffusion XL. Many cards show similar relative performance to SD1.5, but the RTX 5060 actually outperformed the RTX 5060 Ti 8GB in testing. And we've tried it many times. Normally the test only takes a few minutes to run, but it took at 5060 Ti 8GB time Complete the task. I have a driver or application issue that is not working well on NVIDIA's RTX 50 series GPU with 8GB VRAM. Essentially it's like trying to run Stalker 2 with a 4K ultra setting. 5060TI8GB basically doesn't collapse (actual score is 44, but the second white “4” is not visible).
image 1 of 2
ML Commons' MLPERF Client 0.5 test suite provides AI text generation according to various inputs. There are four different tests, all using the llama 2 7b model, and the benchmark measures the time to the first token (the speed at which the response is displayed), and measures the token per second after the first token. These are combined using geometric mean for the overall scores reported here.
AMD, Intel and Nvidia were all ML Commons Partners and were involved in creating and verifying benchmarks, but it appears not to be vendor-independent. AMD and NVIDIA GPUs only have DirectML execution paths, while Intel has both DirectML and OpenVino as options. Intel's ARC GPUS score is much higher for OpenVino than for DirectML.
The 5060 Ti 8GB offers a solid bump of 21% in tokens per second compared to the RX 9060 XT 16GB, with the 8GB and 16GB cards showing relatively similar performance this time. The time to the first token is much better even in Nvidia, taking just 0.22 seconds compared to 0.57 seconds. This is important for real-time conversations, but ultimately, the average token per second rate is a more meaningful indicator in our opinion.
Below are some of the results of a few spec workstation 4.0, but there is an AI inference test that consists of a super-resolution workload running on ResNet50 and a GPU (and potentially NPU). We calculate the geometric average of the four results given in inference per second, which is not the official spec score, but is more useful for our purposes.
The 5060 Ti 8GB offers 25% better performance than the 9060 XT 16GB in the Spec WS4.0 GPU inference test. This is another test that doesn't require 16GB (or 12GB). Both 5060 TI cards have the same performance.
image 1 of 4
For specialized application testing, we'll start with Blender Benchmark 4.3.0, which supports Nvidia Optix, Intel Oneapi, and AMD HIP libraries. These are not necessarily equivalent in terms of optimization levels, but each represents the fastest way to run a blender on a given GPU at the moment. However, I had to use a special preview build of Blender 4.4.0 on my RDNA 4 GPU as the Blender benchmark is not currently running.
Blenders really tend to like Nvidia GPUs, and that applies to the 5060 Ti and 9060 XT. Although gaming performance was often close, Blender means that the RTX 5060 Ti will deliver about twice the performance. This is another test that coincidentally doesn't require more than 8GB of VRAM. Having native Optix support in Blender definitely helps to solve Nvidia.
image 1 of 2
SpecworkStation 4.0 has two other test suites that are interesting when it comes to GPU performance. The first is video transcoding testing using handbrake, scale video engines on various GPUs, and help with content creation tasks. Use the average scores from 4K to 4K to 4K to 1080p. Note that this only evaluates the speed of encoding, not image fidelity.
Both AMD and Nvidia worked to improve their respective video engines with modern architectures, but in the end AMD's 9060 XT offers excellent throughput. It's 39% faster than Nvidia's 5060 TI, but before we declare a true winner here, we need to dig deeper into encoding quality.
image 1 of 7
The final professional app test consists of the Viewport Graphics Suite for SpecworkStation 4.0. This is essentially the same test as SpecViewPerf 2020 and will only be updated to the latest version. (Also, Siemens NX is not part of the suite.) There were seven separate application tests, and each score was combined into an informal overall score using geometric mean.
AMD consumer card drivers tend to be more familiar with these professional applications, which will bring the 9060 XT a huge victory for the 5060 TI. Nvidia offers slightly higher performance with Maya and Solidworks, with the two GPUs tied together with 3DS Max, while other tests are highly favorable for AMD. Overall, AMD's GPU is 48% faster.
These AI and professional tests are, in the end, just one aspect of GPU performance. If you're only concerned about games, it shouldn't have much of an impact on your GPU choice. That applies especially to specialized tests. AI can also be useful for games, but only if you are actually using Blender for 3D modeling, the blender will perform better.
