Nine-year-old Nvidia Tesla V100s fight again: a 4-GPU cluster beats two DGX Sparks at AI inference9 年前英伟达 Tesla V100 老将再战：4 GPU 集群 AI 推理跑赢 2 台 DGX Spark

YouTube channel Jun Yamog published a video on 12 October testing four Tesla V100s mounted on an NVLink carrier board; on the Qwen3.8-Flash-Next benchmark, long-prompt processing was 230% faster than an M5 Ultra and 40% faster than two DGX Sparks. Tesla V100 is Nvidia’s 2017 data-centre GPU based on the Volta architecture, widely used for AI training and inference; the 32GB version has HBM2 memory at 900GB/s. NVLink is Nvidia’s high-speed GPU interconnect for moving data directly between GPUs, cutting the interconnect bottleneck compared with PCIe in multi-GPU servers and AI clusters. The carrier board offers 300GB/s of bandwidth and six NVLink 2.0 connections. To house four separate GPUs the tester modified an Nvidia A100 heatsink to dissipate roughly 300W per V100 and connected the board over PCIe x8. The host ran a 4-bit quantised Qwen3.8-Flash-Next and compared it with two DGX Sparks with 128GB of unified LPDDR5X memory and a Mac Studio with an M5 Ultra. In long-prompt processing the four V100s were 40% faster than the DGX Spark pair and 230% faster than the M5 Ultra; the tester notes the huge gap with Apple Silicon may partly reflect insufficient software optimisation. In short-prompt processing the V100 cluster kept a 40% lead over the DGX Sparks and was 163% faster than the M5 Ultra, showing older data-centre GPUs still deliver high throughput on certain inference workloads. In short-prompt token generation the four V100s hit 100.6 tok/s, 121% faster than two DGX Sparks; the M5 Ultra managed 44.5 tok/s.