The short version: A modified system built around four older NVIDIA Tesla V100 GPUs reportedly outperformed newer compact AI machines in selected inference benchmarks published in October 2026. The comparison is interesting for people running local language models, but it is not proof that a nine-year-old GPU is always faster, cheaper or more efficient than new hardware.
What is the NVIDIA Tesla V100?
NVIDIA introduced the Tesla V100 using the Volta architecture in 2017. Versions with substantial high-bandwidth memory were designed for data-centre workloads rather than ordinary desktop gaming. Because enterprise hardware eventually reaches the second-hand market, technically experienced users can sometimes assemble compute setups at lower upfront equipment prices than brand-new specialist machines.
The reported experiment used four V100 GPUs connected through an NVLink carrier board, along with cooling modifications. The October 12, 2026 Wccftech report describes selective gains compared with Apple’s M5 Ultra and NVIDIA’s DGX Spark. Some result descriptions use different percentage baselines, so headline claims of “up to 230%” should not be treated as a universal gain across models or tasks.
Why might multiple old GPUs beat one newer device?
Memory capacity and bandwidth: Large model inference can be limited by moving model weights, not merely raw arithmetic. Several cards together can provide additional capacity if the workload and software can effectively distribute weights.
Interconnect: NVLink can reduce some device-to-device transfer bottlenecks. However, it does not make four GPUs behave like one perfectly unified pool of memory; the efficiency depends on the framework and parallelism strategy.
Benchmark choice: Prefill throughput, time to first token and output tokens per second are different metrics. A setup tuned to one model size or context length can appear outstanding in a narrow test and weaker elsewhere.
The hidden costs of a bargain inference cluster
Second-hand server hardware can require unusual power supplies, specialist cooling, loud fans and complicated drivers. Energy usage matters over months of operation. There is also failure risk, warranty uncertainty and substantial setup time. A compact new workstation may be slower in a particular test but easier to maintain, quieter and more practical for a small office.
Teams should compare total cost of ownership, not the price of a GPU alone. Include all cards, carrier boards, cooling, electricity, housing, software maintenance and operator time. Also check compatibility with inference frameworks and desired model precision.
How to compare AI inference systems fairly
- Choose the same model revision and quantization for every device.
- Use identical prompt lengths, generation lengths and batch sizes.
- Record output tokens per second, first-token latency and memory use.
- Measure wall-plug power during real workloads rather than relying only on rated power.
- Repeat tests and disclose software versions, cooling modifications and failures.
For teams comparing local deployment with cloud alternatives, see BCC’s guide to private LLM developer workflows. For a broader technology context, read our report on hardware supply-chain constraints.
Frequently asked questions
Is Tesla V100 still good for AI inference in 2026?
It can remain useful for specific models and workloads, especially when purchased cheaply and run by skilled operators. That is not a recommendation for every buyer.
Is four V100 GPUs always faster than DGX Spark?
No. The reported results apply to selected tests, configurations and metrics, not all AI workloads.
Should beginners build this system?
Not without planning for server-grade power, cooling, interconnects, software compatibility and maintenance. A supported modern system may be a better first deployment.
Source and methodology note
Wccftech report of the modified V100 benchmarks (12 October 2026). BCC has not independently reproduced the benchmark, so comparative performance claims are attributed to the reported test rather than presented as independently verified hardware specifications.