Chinese LLM DeepSeek wrecked havoc in the US tech market, effectively wiping out trillions of dollars from the stock market. Despite being trained on slightly older Nvidia hardware, it works surprisingly well on one of AMD's consumer-grade offerings: the Radeon RX 7900 XTX. David McAfee, GM of AMD's Radeon division, posted some benchmark results on X.
Depending on the number of parameters and model used, the Radeon RX 7900 XTX is anywhere between 13% (7 billion parameters) and 2% (14 billion parameters) faster. After that, the RDNA 3 flagship begins to falter and falls behind the RTX 4090 with 32 billion parameters. AMD even compared it with the GeForce RTX 4080 Super, where the 7900 XTX has a 34% performance lead.
AMD has also provided detailed instructions on how to run DeepSeek locally on your machine. But, the Radeon RX 7900 XTX mentioned above caps out at 32 billion parameters, while the Strix Halo Ryzen AI Max 395 Plus with 128 GB RAM, can support up to 72 billion. Or, if you have about $6,000 to spare, Matthew Carrigan has figured out a way to run the entire model locally on a machine with two AMD Epyc CPUs and 768 GB of RAM.