XPerf Inc. announced its deployment with MiTAC Computing: ClusterReady™ demonstrated that the AMD Instinct™ MI325X AI air-cooled racks — featuring MiTAC Computing's G8825Z5 servers — achieved performance 20% higher than AMD's publicly reported FLUX.1-dev benchmark results under the tested configuration, and produced the first reported Llama 3.1 8B pre-training performance results on MI325X. Across the full 4-node rack, XPerf's network optimization delivered a 10× improvement in multi-node communication bandwidth, enabling 95.8% scaling efficiency on Llama 3.1 8B pre-training. Full MLPerf® Training V6.0 benchmark results have been submitted.
Conventional L11/L12 rack tests verify integration and functional readiness but cannot replicate live sustained AI workloads — a rack that passes every standard test can still fail in production, eroding the ROI operators depend on. ClusterReady™ stress-tests racks using a combination of inference and training AI workloads alongside XPerf's proprietary test suite — exercising GPUs, memory, interconnects, network fabric, and thermal systems under conditions that mirror live production. ClusterReady™ measures tokens per second output under sustained load, giving a concrete, revenue-linked performance baseline for every rack. When issues are detected, XPerf's proprietary algorithms identify the root cause and, uniquely among validation tools, prescribe specific remediation or optimization steps, reducing manual engineering effort.
Key features
- 10× Faster Validation — XPerf's proprietary algorithms and automated workflow compress full rack validation from weeks to hours or days, so racks reach revenue-generating production in a fraction of the time.
- Stress Test Suite — combines inference and training AI workloads with XPerf's proprietary test suite to stress GPUs, memory, interconnects, network fabric, and thermal systems, exposing hardware and firmware issues that conventional methods miss.
- Root Cause Analysis with Prescribed Solutions — the only validation tool that identifies the root cause of detected issues and prescribes specific remediation and optimization steps, cutting engineering triage time and getting racks ready faster.
- Revenue-Linked Performance Baseline — every rack ships with a ClusterReady™ validation report documenting tokens per second output across inference and training workloads, providing a performance baseline for revenue generation.
“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.”
“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.”
Availability
ClusterReady™ is commercially available and designed for rapid deployment in supported environments. To learn more or schedule a ClusterReady™ validation demo, visit xperf.ai/clusterready.