A recent blog post from a Chinese developer provides a practical comparison of fine-tuning performance on two consumer GPUs: the RTX 5090 with 32GB VRAM and the RTX 4090D with 24GB. The test involved training Qwen3.5-9B using QLoRA, a method that combines 4-bit quantization with LoRA adapters. The reported result is a 9% improvement in adoption rate (a metric likely related to model output quality or acceptance) on the RTX 5090. The author explains this gap by pointing to the information bottleneck created by 4-bit quantization: compressing 16-bit weights to 4-bit inevitably loses information, even with NF4 quantization. While the RTX 5090's larger memory allows for higher precision or larger batch sizes, the post suggests that the quantization itself is a limiting factor. For ML engineers and teams evaluating hardware for fine-tuning, this provides a useful data point, though the specific metric and methodology would need further scrutiny. The post also touches on the broader trade-off between memory efficiency and model quality in QLoRA-style training.
A Chinese blog reports that QLoRA training of Qwen3.5-9B on RTX 5090 (32GB) achieves a 9% higher adoption rate than on RTX 4090D (24GB), attributing the gap to 4-bit quantization information loss.