Old Nvidia V100 GPU Cluster Competes with Newer Systems in AI Tests
A setup with four older Nvidia Tesla V100 graphics cards connected via NVLink demonstrated significant performance in AI tasks, partially outperforming even newer systems.
A test conducted by YouTube user Jun Yamog shows that a configuration featuring four Nvidia Tesla V100 graphics cards achieves notable performance in AI inference tasks. Despite the V100 cards being released in 2017, they managed to outperform even newer and more powerful systems in certain tests.
The benchmark utilized the 4-bit quantized Qwen3.8-Flash-Next model. The cluster of four V100 cards, connected via NVLink high-speed interconnect technology, processed long prompts 40 percent faster than two newer DGX Spark systems. Compared to Apple's M5 Ultra chip, the performance was up to 230 percent better.
The tests also revealed that the V100 cluster was 40 percent faster in short prompt processing and 163 percent faster compared to the M5 Ultra. In short token generation, the V100 configuration achieved 121 percent higher speed than the DGX Spark system.
These results suggest that older datacenter GPUs can still offer substantial throughput for specific AI inference workloads. The NVLink technology, enabling direct GPU-to-GPU communication, appears crucial in maintaining competitive performance even with legacy hardware.