XPerf Inc. has announced the deployment of its ClusterReady™ rack validation and optimization software with MiTAC Computing, demonstrating that air-cooled racks equipped with MiTAC’s G8825Z5 servers and AMD Instinct™ MI325X GPUs achieved performance 20% higher than AMD’s publicly reported FLUX.1-dev benchmark results. Additionally, the deployment produced the first reported Llama 3.1 8B pre-training performance results on the MI325X, enabling a 95.8% scaling efficiency across the 4-node rack due to a 10× improvement in multi-node communication bandwidth. While conventional L11/L12 tests fail to replicate sustained AI workloads, ClusterReady™ stress-tests infrastructure using an advanced combination of inference and training workloads, reducing full rack validation time from weeks to hours or days through automated root cause analysis and prescribed remediation steps.
Also Read: NVIDIA and Microsoft Partner to Reinvent Windows PCs for AI Agents via RTX Spark Superchip
Highlighting this success, Raymond Huang, GM, MiTAC Computing Technology Corp., USA stated, “XPerf delivered excellent results on cluster-level validation and benchmarking within days work that typically takes weeks. Our AMD MI325X AI air-cooled rack validated seamlessly and achieved 20% stronger performance on industry-relevant AI workloads than the standard AMD MI325X rack, highlighting MiTAC’s engineering expertise in building high-performance AI infrastructure.” Alex Carter, Founder & CEO, XPerf Inc. added, “The gap between assembling a GPU rack and knowing its performance has cost this industry too much for too long. Partnering with MiTAC, a leading global AI server manufacturer, we validated and optimized a MI325X rack in days and proved it outperforms AMD’s published benchmarks by 20%. The same speed applied to the validation of a MI350X rack. That’s the standard ClusterReady™ sets.”
