AMD has released guidance on operating the DeepSeek R1 AI model using its AI-enhanced Ryzen AI and Radeon hardware, simplifying the process for users to deploy this advanced chain-of-thought model on their personal computers. The R1 model is compatible with several large language models (LLMs) and can be run on the RX 7000 series desktop GPUs and specific Ryzen CPUs equipped with XDNA NPUs, provided that users install the optional Adrenalin 25.1.1 driver.
The instructions provided by AMD include all the necessary steps for setting up DeepSeek R1 on compatible local systems. AMD users will primarily utilize the LM Studio’s one-click installer designed specifically for Ryzen AI to install R1. AMD also details how to optimize the application for their hardware, providing a list of the maximum LLM parameters that their devices can support.
It’s reported that DeepSeek R1 has been recently refined into smaller, yet highly effective models that are manageable on consumer-grade hardware. For a bit of perspective, the original DeepSeek-V3 model was trained using a massive cluster of 2,048 Nvidia H800 GPUs.
The capacity for LLM parameters is determined by the available memory. Models such as the RX 7600 XT, 7700 XT, 7800 XT, 7900 GRE, and 7900 XT can handle up to “DeepSeek-R1-Distill-Qwen-14B”. The top-tier RX 7900 XTX is capable of supporting up to “DeepSeek-R1-Distill-Qwen-32B”. Meanwhile, the RX 7600, equipped with 8GB of VRAM, can manage up to “DeepSeek-R1-Distill-Llama-8B”.
In the realm of mobile APUs, the Ryzen 8040 and 7040 series come with 32GB of RAM, and the Ryzen AI HX 370 and 365 have 24GB and 32GB of RAM respectively, supporting up to “DeepSeek-R1-Distill-Llama-14B”. The Ryzen AI Max+ 395 can handle up to “DeepSeek-R1-Distill-Llama-70B”, but this is contingent on having memory capacities of either 128GB or 64GB; with 32GB, the support is limited to “DeepSeek-R1-Distill-Qwen-32B”.
The innovative DeepSeek R1 AI model has been making significant waves globally, thanks to its computing cost, which is 11 times lower than that of the most advanced models currently available. Just two days ago, it was instrumental in causing a record-breaking $589 billion drop in Nvidia’s market cap. This model achieves its remarkable 11X efficiency gain through extreme optimization, utilizing Nvidia’s assembly-like Parallel Thread Execution (PTX) programming to boost performance predominantly.
While Nvidia and AMD are the main players in supporting R1, Huawei has also integrated DeepSeek support into its Ascend AI GPUs, allowing efficient AI performance on Chinese-made hardware.
