Physix Frontier · News Briefing Card (IT Home · Oct 13, 2026)

4 Nine-Year-Old Tesla V100s Beat 2 DGX Sparks in AI Inference

KEY FACTS

YouTube channel Jun Yamog published a test video on October 12.
The test ran 4 Tesla V100s on an NVLink baseboard.
In long-prompt processing, the cluster was 40% faster than 2 DGX Sparks.
In long-prompt processing, the cluster was 230% faster than an M5 Ultra.
In short-prompt generation, the 4 V100s reached 100.6 tok/s, 121% faster than a DGX Spark.

KEY DATA

40%Long-prompt processing speedup vs DGX Spark

230%Long-prompt processing speedup vs M5 Ultra

100.6 tok/sShort-prompt token generation speed

121%Short-prompt generation speedup vs DGX Spark

PHYSIX OBSERVATION

Old cards clustered over NVLink can still outpace new devices in inference throughput, showing that interconnect bandwidth and the software stack matter more than single-card generation. For budget-constrained teams, a used V100 cluster is a short-term cost-cutting option, but power draw and ecosystem support are hidden costs. Apple devices lag by a wide margin, a clear signal of insufficient software optimization.

Source: IT Home report