Older NVIDIA gpus outperform newer systems in AI inference tests

A YouTube channel tested four Tesla V100 GPUs, installed on an NVLink board, and found they outperformed two DGX Spark systems and an M5 Ultra chip in AI inference benchmarks. Launched in 2017, the V100s ran the Qwen3.8-Flash-Next model with 4-bit quantization, achieving a 40% speed boost over DGX Spark and a 230% increase over the M5 Ultra. The setup used a modified A100 cooler to manage power consumption and connected the V100s via PCIe x8. The results suggest that software optimization may play a significant role in performance differences between Apple Silicon and other platforms. In short prompt processing, the V100 cluster maintained a 40% edge over DGX Spark and a 163% lead over M5 Ultra. The V100s also achieved 100.6 tokens per second in token generation tests, outperforming DGX Spark by 121%. The test underscores the continued relevance of legacy GPU technology in AI inference, even as newer systems like DGX Spark and Apple's M5 Ultra gain traction.