Report: NVIDIA P106-based system scores over 30 tokens per second
A custom rig with nine NVIDIA P106 cards reportedly achieves high performance on large language models.
According to Guru3D, a custom-built system featuring nine NVIDIA P106 cards has been reported to deliver over 30 tokens per second on the Qwen3.6 35B-A3B-class model. Each of the cards is equipped with 6 GB of VRAM.
The buyer claims that the rig's total 54 GB of VRAM is sufficient for running an 8-bit quantized 27B-class model and can handle up to approximately 70 billion parameters with 4- or 5-bit quantization. Despite lacking Tensor Cores, all nine cards were functional in the system, which only required an SSD to be fully operational.
The reported setup is constrained by PCIe 3.0 x1 connections between the GPUs, limiting inter-GPU bandwidth. However, the buyer noted that the rig's high power draw is offset by free electricity at their accommodation.
Key facts
- - Nine NVIDIA P106 cards with 6 GB VRAM each
- - Over 30 tokens per second on Qwen3.6 35B-A3B-class model
- - 54 GB total VRAM for large language models
- - PCIe 3.0 x1 connections between GPUs
Sources
- Guru3DRetired nine-GPU mining rig bought for $35 provides 54GB VRAM for local LLM inference