9-Year-Old NVIDIA Tesla V100 Cluster Outperforms Newer Hardware in AI Inference
A YouTuber demonstrated that a custom cluster featuring four 2017-era NVIDIA Tesla V100 GPUs connected via NVLink outperformed hardware like the Mac Studio M5 Ultra and DGX Spark in long-prompt AI inference benchmarks. This experiment highlights how high-speed multi-GPU interconnect bandwidth like NVLink can keep older data center hardware surprisingly competitive in specific large language model inference workloads. The setup utilized an NVLink carrier board providing 300GB/s bandwidth and modified NVIDIA A100 heatsinks to manage the 300W power draw of each V100 card, running a 4-bit quantized Qwen3.8-Flash-Next model.
## BACKGROUND
The NVIDIA Tesla V100 is a Volta-architecture data center GPU released in 2017 that utilizes HBM2 memory and provides high memory bandwidth for AI tasks. NVLink is a proprietary high-speed interconnect technology developed by NVIDIA to enable direct, high-throughput communication between multiple GPUs without relying solely on PCIe bottlenecks.