For the past weeks or so, no AI news has seemed complete without mentioning DeepSeek. The new LLM AI model hasn't only been a concern among Western AI companies, but domestic rivals like Alibaba have also started to put up strong competition against it. While the hype around DeepSeek's low cost compared to other AI models is well known, its low computational requirements are another area where it seems to be making strides, as an average consumer can achieve adequate performance when running the model.

This comes after AMD shared DeepSeek’s R1 AI inference benchmarks, comparing the flagship Radeon RX 7900 XTX GPU with NVIDIA’s counterpart from the RTX 40 series, showing superior performance across multiple models. On top of that, AMD also seems to be quickly pushing out support for DeepSeek's R1 LLM models, as consumer GPUs for AI workloads have worked well for several individuals and it looks like the Red Team wants to capitalize on this opportunity first.

RDNA 3-based RX 7900 XTX outperforms NVIDIA’s last-gen flagship GPUs

As per the graphic shared by AMD, the Radeon RX 7900 XTX handles AI inference workloads much better than the RTX 4080 Super and even outperforms the RTX 4090 in the majority of tasks. Before diving into the benchmarks, it’s important to understand what the conducted tests represent. DeepSeek R1 is a set of AI models designed for various tasks, available in different sizes, such as Distill Qwen 7B and Distill Llama 8B. The number represents the billions of parameters in the model—the larger the number, the more complex and demanding the model is.

With that out of the way, in performance comparisons, the RX 7900 XTX showed significant gains over the RTX 4080 Super in nearly all tests, especially in the Distill Qwen 7B test, where it reached up to 134% of the 4080 Super's performance—making it about 34% faster in AI inference speeds. Against the more powerful RTX 4090, the RX 7900 XTX still managed to come out ahead in nearly all cases. The most notable lead was in the Distill Qwen 7B test, where it reached up to 113% of the RTX 4090's performance, showing a 13% advantage. However, in the Distill Qwen 32B test, the RX 7900 XTX dropped to 96% of the RTX 4090's performance, falling slightly behind when handling a much larger and more demanding model.

Guide to running R1 on your local AMD machines

DeepSeek’s AMD GPU benchmarks weren't the only highlight from the Red Team, as they also released an extensive guide on running DeepSeek R1 distillations on your local AMD GPU. Additionally, AMD published a YouTube tutorial that walks through the same steps individually, catering to those who prefer a visual guide over written instructions.

That said, with such strong AI performance on an RDNA 3 GPU, many users are now wondering how big of a jump we can expect from the upcoming RX 9070 series GPUs, which will be based on the RDNA 4 architecture. With NVIDIA's RTX 50 series also heavily focused on AI performance, the competition between next-gen GPUs is shaping up to be tough. However, with AMD delaying its launch to March, there’s still some time to go before we’ll find out.

About the Author

Hassam Nasir

Hassam boasts over seven years of professional experience as a dedicated PC hardware reviewer and writer.