Library developers GNU MP (GMP) have issued a warning to users regarding the incident involving the failure of AMD Ryzen 9000 series processors based on the Zen 5 microarchitecture during testing of the GMP library. It is believed that the damage to the processors was caused by overheating, which occurred due to the intensive performance of GMP functions. Until the issue is resolved, users are advised against actively using GMP on systems with AMD Zen 5 CPUs.
The concerns stem from the repeated failure of the AMD Ryzen 9950x CPU (AMD Zen5 X16 4300MHz) in the test infrastructure. The initial failure occurred in February 2025, leading to a replacement under warranty. However, at the end of August, a second CPU failed in a system with a different power supply and motherboard, displaying the same signs of overheating. Notably, a bleached area of approximately 1 centimeter was observed on the chip.
Initial findings suggest that the GMP library is utilizing Mulx cycles excessively on AMD Zen 5 systems, resulting in significantly higher energy consumption than specified and overwhelming the cooling system’s capacity. Both instances of damage occurred during tests involving manual assembler cycles with a single Mulx instruction, intensifying the load on the processors.
The systems involved in the tests utilized ASUS Prime B650M-K and Asus Prime B650M-A WiFi II motherboards, along with Corsair SF450 and Corsair RM650 power supplies. In both configurations, a NOCTUA NH-U9S cooling system was employed, positioned differently based on manufacturer recommendations to optimize cooling efficiency.
It has been observed that in the second instance of failure, the thermal paste was unevenly distributed, potentially contributing to the problem. The Ryzen 9950X CPU requires a thermal design power of 170W, while the cooling system used provided thermal dissipation up to 165W. The shortfall of 5W was compensated for by additional case fans and a low ambient temperature of around 20°C. Both processors failed after several months of operation under high load, despite another server in the same configuration performing effectively in the project infrastructure for an extended period.