AMD's flagship graphics card, the RX 7900 XTX, has demonstrated superior performance over NVIDIA's RTX 4090 in recent DeepSeek AI model testing, according to benchmarks released by AMD.
The tests, shared by David McAfee on X, show the RDNA3-based GPU outperforming the RTX 4090 by up to 13% when running DeepSeek R1 Distill Qwen 7B. The RX 7900 XTX also maintained its lead in two other configurations, showing an 11% advantage with Distill Llama 8B and a 2% edge using Distill Qwen 14B.
When compared to the RTX 4080 Super, AMD's card displayed even more impressive results, achieving up to 34% better performance with DeepSeek R1 Distill Qwen 7B. The performance gap remained substantial at 27% with Distill Llama 8B and 22% with Distill Qwen 14B.
The RX 7900 XTX features 192 AI accelerators and supports BF16 and INT8 operations, making it well-suited for AI workloads despite not being primarily marketed as an AI processor. AMD has recently released detailed instructions for running DeepSeek R1 on compatible consumer hardware, including the RX 7900 XTX.
While these results appear promising for AMD, they should be interpreted cautiously as they come from AMD's internal testing. The specific configuration of the NVIDIA GPUs during testing remains unclear, and not all AI workloads fully utilize a GPU's computational capabilities.
DeepSeek R1, the AI model used in testing, has gained attention for offering performance comparable to leading Western AI models while requiring less computational power. The model employs various hardware optimizations, including NVIDIA's PTX programming language, to achieve speeds up to 11 times faster than competitors.