AMD Announces Support for DeepSeek R1 Distillate Models on Radeon RX and Ryzen AI GPUs | Cpu hardware list and functions | Nvidia hardware | Hardware and software of computer | Turtles AI

AMD Announces Support for DeepSeek R1 Distillate Models on Radeon RX and Ryzen AI GPUs
New Ryzen AI Strix Halo Processors and Radeon RX 7000 Series Deliver Advanced Performance for AI Model Acceleration with DeepSeek R1
Isabella V29 January 2025

 

AMD today released a guide to running DeepSeek R1 distilled reasoning models on Radeon RX graphics cards and Ryzen AI processors. The new configurations, including Ryzen AI Max “Strix Halo” processors and RX 7000 graphics cards, promise advanced performance in distilling complex models.

Key Points:

  • The new Ryzen AI Max “Strix Halo” processors are configured with up to 128GB of LPCAMM2 memory.
  • Radeon RX 7000 graphics cards, especially the RX 7900 XTX, are optimized to accelerate the distillation of DeepSeek R1 models.
  • The guide distinguishes performance based on memory and chip model, recommending different configurations for each scenario.
  • The tests show a performance advantage of the RX 7900 XTX over NVIDIA GeForce RTX, with significant improvements in various distillation models.

AMD recently released a detailed user guide on how to optimize the execution of DeepSeek R1 distilled reasoning models on its new processor and graphics card lines. The main protagonists of this guide are the new Ryzen AI Max "Strix Halo" processors, configured with LPCAMM2 memory in 32, 64, and 128 GB capacities, but excluding 16 GB options. This choice is intended to favor superior performance in notebooks powered by these processors, which will be able to run complex models like the DeepSeek-R1-Distill-Llama with a whopping 70 billion parameters in the 64 GB and 128 GB configurations. The 32GB version, on the other hand, is designed to run patterns like DeepSeek-R1-Distill-Qwen-32B, while the smaller configurations, with the "Strix Point" processors based on RDNA 3.5 iGPU and NPU, will support smaller patterns like DeepSeek-R1-Distill-Qwen-14B and DeepSeek-R1-Distill-Llama-14B. AMD also recommends running these patterns using Q4 KM quantization to maximize performance.

Moving on to graphics solutions, the company indicated that the Radeon RX 7000 series cards, which integrate the new RDNA 3 architecture with dedicated AI accelerators, are the most suitable for running distilled patterns. In particular, the Radeon RX 7900 XTX stands out as the best choice for distilling models like DeepSeek-R1-Distill-Qwen-32B, while the versions with memory between 12 and 20 GB (RX 7600 XT, RX 7700 XT, RX 7800 XT, RX 7900 GRE and RX 7900 XT) stop at the lighter models like DeepSeek-R1-Distill-Qwen-14B. Even the Radeon RX 7600, with 8 GB of memory, is limited to managing smaller models like DeepSeek-R1-Distill-Llama-8B.

In terms of performance, AMD provided real-world test results, pitting the RX 7900 XTX against NVIDIA’s GeForce RTX 4080 SUPER and RTX 4090. Compared to the RTX 4080 SUPER, the RX 7900 XTX showed significant improvement, with up to 34% better performance in the DeepSeek-R1-Distill-Qwen-7B model and up to 27% better performance in the DeepSeek-R1-Distill-Llama-8B model. The comparison with the RTX 4090 is equally interesting, with the RX 7900 XTX prevailing in three out of four tests performed. In particular, it recorded 13% better performance with DeepSeek-R1-Distill-Qwen-7B and 11% better performance with DeepSeek-R1-Distill-Llama-8B. However, in the case of the DeepSeek-R1-Distill-Qwen-32B model, the RX 7900 XTX was only 4% behind the RTX 4090.

AMD emphasized that to take full advantage of all this performance, you need to have LMStudio software version 0.3.8 or later and Radeon Software Adrenalin 25.1.1 beta drivers or higher.

Video