XPerf has launched from stealth, announcing a software platform designed to increase GPU and application-specific integrated circuit (ASIC) utilization for AI data centers operating multi-generation, multi-vendor environments. XPerf claims its AI-native optimization platform can improve GPU utilization by at least 20 percent, directly addressing operational inefficiencies in complex AI and high-performance computing clusters.
XPerf’s platform targets the persistent challenge of low hardware utilization observed in modern AI infrastructure. The company cites Uptime Intelligence, which reports that GPUs are operational only 80 percent of the time, with typical models reaching just 35 to 45 percent of maximum compute performance. Coreweave reported the highest Model FLOPS Utilization on Nvidia H100 GPUs at 52 percent in large-scale training benchmarks. XPerf’s solution uses advanced algorithms to dynamically optimize resources across heterogeneous hardware, workload types, and energy profiles, adapting to evolving AI workflows, including pre-training, post-training, inference, and other emerging tasks.
According to XPerf, its full-stack software integrates with diverse GPU, ASIC, and networking deployments, automating workload and resource management to maximize performance and energy efficiency. The team, co-founded by engineers with prior Intel Gaudi experience, has built the platform on top of proprietary hardware monitoring and management technology.
XPerf is partnering with Infraeo, a company specializing in high-speed copper and optical interconnects for AI clusters and data centers, and will demonstrate the optimization solution at booth 5412 during SC25. Infraeo, an early investor in XPerf, claims its connectivity solutions are designed to operate at scale in demanding AI and high-performance computing environments.
XPerf reports that its platform is positioned for large-scale data centers developing or deploying GPU clusters and custom ASICs for AI workloads. XPerf is a member of the Nvidia Inception program, the Cloud Native Computing Foundation, and gBETA accelerator program.
Dr. Molly Guo, CTO and Co-Founder of XPerf, stated, “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,”
Dr. Rakesh Sambaraju, Executive Vice President of Infraeo and early investor in XPerf Inc., noted, “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,”
Source: XPerf
