High-Performance Computing Enthusiasts Showcase Massive Multi-GPU Configurations
A growing trend of high-end, custom-built hardware rigs designed for large language model (LLM) inference is emerging within enthusiast communities. Recent reports from users on Reddit highlight the increasing complexity of these systems, which utilize multiple GPUs to manage the high VRAM requirements of modern AI models.
One notable configuration involves an upgraded system featuring eight R9700 GPUs. The build utilizes an Asrock WRX 90 WS EVO motherboard paired with a Threadripper 9955WX processor and 128GB of DDR5 5600 RAM. To manage the significant power requirements, the setup employs a dual power supply configuration consisting of a 2200W Seasonic and a 1600W Seasonic unit.
The user reported that the expansion from four to eight GPUs was intended to accommodate larger upcoming models. In preliminary testing, the system demonstrated varying performance metrics depending on the model and inference engine used. According to the user, running QFN with radiance provided a high-quality experience, while the GLM 5.3 Flash model achieved approximately 30 tokens per second (t/s) during decoding at 4-bit quantization, with a prefill rate of roughly 2000 t/s.
Performance Benchmarks and Hardware Efficiency
Discussions surrounding these builds have also touched upon the efficiency of older enterprise-grade hardware. One user noted that a setup utilizing eight Radeon Pro V620 GPUs—which provide 256 GB of VRAM—offered a cost-effective alternative to newer hardware. While the V620s have lower bandwidth (512 GB/s), they remain functional for specific workloads.
Performance gains were noted when using custom software forks. For instance, one user reported that using a custom vLLM fork allowed the Qwen3.8-Flash-Next model to reach decoding speeds of 70-100 t/s and prefill speeds between 4000-5000 t/s.
Regarding power consumption, the hardware’s idle state is relatively low, with each GPU consuming approximately 8 watts. However, power draw increases significantly during the prefill stage when all cards are active simultaneously.
Software-Induced GPU Utilization Concerns
While high-performance builds focus on maximizing utility, other users have reported issues regarding unintended GPU resource consumption. A report on Reddit highlighted a potential issue with MuseHub’s background service (Muse.Service), which allegedly causes high GPU utilization on idle systems.
The user reported that the service appears to pin the GPU at 100% utilization in a rhythmic loop, cycling between 20 to 40 seconds of full load and approximately six seconds of inactivity. The user expressed concern that such unprovoked workloads could lead to hardware degradation, particularly in components already experiencing surface oxidation.
At this time, these reports of MuseHub’s behavior remain unverified by independent technical audits. Users experiencing similar rhythmic spikes in GPU usage via Task Manager have been advised to investigate background services or manually disable the service via command line.