Vectris Labs Claims Up to 73% Boost in GPU Inference Throughput

The company’s internal testing on NVIDIA H100, H200, and B200 GPUs revealed significant performance gains, with throughput increasing by 30% to 73% depending on the specific hardware. Beyond raw output, the Waveform control plane demonstrated a 51% to 56% reduction in energy consumption and accelerated workload completion by 22% to 42%. These improvements rely on identifying structural inefficiencies during inference execution and reorganizing them in real time.

Vinod Tipparaju, co-founder and CTO of Vectris Labs, noted that the team did not force this structure onto the computation, but rather discovered it as an inherent, measurable property of AI workloads. By formalizing this as Compute Yield, the company aims to provide a new economic metric for data center operators facing power and capital constraints. While current results are workload-specific and have yet to be independently reproduced in large-scale production environments, the technology has already been validated on Intel and AMD silicon using the MLPerf LoadGen benchmark.

Vectris plans to move from these initial demonstrations to a limited launch of Waveform on October 1, 2026. The firm, which was incubated by Thumos Capital, intends for the platform to eventually extend beyond GPU inference into networking, memory, and thermal systems to optimize the broader AI infrastructure stack.