TensorBench Open-Source Benchmark to Measure What AI Compute Actually Costs
Startup Tensor Machines released an open-source benchmark, TensorBench, that measures useful AI output per watt of GPU power — a building block for the verifiable efficiency standards AI sustainability work still lacks.
Tensor Machines, a startup building physics-based models for AI infrastructure operators, has released TensorBench, an open-source benchmark on GitHub that ties GPU power draw and thermal behavior to how much useful AI output a data center actually gets — rather than leaning on vendor spec sheets or Power Usage Effectiveness (PUE), the industry-standard metric that measures facility overhead but says nothing about the AI work being done with that power. In early tests on NVIDIA GPUs, the same hardware produced up to 15% more tokens per second depending on workload, and raising the power cap drew about 36% more electricity for only 21% more throughput — inefficiency that's effectively invisible to today's standard metrics.
The company, which also announced $1.5 million in pre-seed funding led by Omni VC, presented the tool at the NeoCloud Summit in San Francisco this month. Open, independently reproducible measurement is a precondition for any real sustainability standard in AI, and a benchmark published on GitHub rather than locked inside a vendor dashboard is exactly the kind of building block AIforE wants to see more of: a common yardstick that lets operators, auditors and certifiers compare efficiency claims rather than take a cloud provider's word for its own numbers. It's a first step, not a certification — TensorBench still stops short of tying output to carbon or water figures, which remains the next gap for groups like AIforE to help close.
Source: PR Newswire
