4 old NVIDIA V100 GPUs beat M5 Ultra in long prompt speed test

A YouTube channel called Jun Yamog tested four NVIDIA Tesla V100 data center GPUs from 2017 on an NVLink board, comparing them with the M5 Ultra and two DGX Spark units.

Using a 4-bit quantized Qwen3.8-Flash-Next model, the V100 setup was 230% faster than the M5 Ultra and 40% faster than two DGX Sparks on long prompts. On short prompts, the leads were 163% and 40%.

The V100 uses Volta architecture, with the 32GB HBM2 version offering 900GB/s memory bandwidth. The NVLink board gives 300GB/s of interconnect, easing multi-GPU bottlenecks. The tester thinks part of the gap with Apple Silicon comes from software optimization.

In short prompt token generation, the V100 cluster hit 100.6 tok/s, while the M5 Ultra reached 44.5 tok/s. Old data center GPUs can still deliver solid throughput for some AI workloads.