XPerf Inc., an AI infrastructure startup specializing in GPU and ASIC utilization optimization, today emerged from stealth mode and announced plans to accelerate the development and commercialization of the full-stack GPU utilization optimization software platform.
The utilization gap
As AI infrastructure evolves from training-dominated workloads to a mix of pre-training, post-training, inference, and emerging applications, its complexity continues to grow. The deployment of multi-generation, multi-vendor GPUs, ASICs, and diverse networking fabrics makes GPU utilization optimization increasingly challenging, requiring technologies that dynamically balance hardware capabilities and workloads.
This inefficiency erodes the return on investment, increases infrastructure OPEX, delays innovation and applications time-to-market, and inflates the carbon footprint of AI operations.
An AI-native approach
Co-founded by veteran engineers with extensive experience in deploying AI clusters with thousands of accelerators and hundreds of deployment cases, the XPerf team has been working on a novel AI-native approach with advanced algorithms to address the multi-dimensional challenges of hardware and network inhomogeneity, energy consumption, and workloads.
“With our unique AI-native method, XPerf Inc.'s technology can boost the AI data centers' GPU utilization by at least 20% while meeting the dynamic workload demands.”
“As AI data center deployments are designed towards inferencing, optimizing utilization and cost becomes critical. The challenge that XPerf is addressing and the solution it is developing will be extremely valuable for our customers who are either deploying large-scale GPU clusters or developing custom ASICs for AI applications.”
Showcasing at SC25
XPerf Inc. showcased its technology with partner Infraeo Inc., a cutting-edge AI cluster copper and optical connectivity interconnect solutions company, at SC25 (booth #5412).