XPerf™ Brings Its AI Cluster Lifecycle to SC26: ClusterBeacon™ Now Available, ClusterMarshal™ Opens for Pilots

XPerf’s ClusterBeacon forecasts GPU failures 10–30 minutes ahead, and ClusterMarshal suggests where, when and how each job runs. See both at SC26, booth 5617.

ROUND ROCK, TX, UNITED STATES, November 10, 2026 /EINPresswire.com/ — XPerf Inc., an AI infrastructure software company building the control plane for heterogeneous data centers, today announced two product milestones ahead of SC26: ClusterBeacon™, an AI cluster failure prevention and resolution software, is now available for commercial deployment, while ClusterMarshal™, a scheduler for AI workloads on heterogeneous hardware, is opening for pilot deployments. Together with ClusterReady™, XPerf’s AI rack validation platform, the three products address the full AI cluster lifecycle.

XPerf will demonstrate its portfolio at SC26 in Chicago, November 17–19, at booth 5617, shared with Infraeo.

More output from the GPUs already installed.
With power limiting how quickly new data center capacity can be built, more output has to come from the GPUs already in service. In one published study of a 4,704-GPU cluster, GPUs read 97–99% busy while doing useful compute only about 40% of the time (Hu et al., NSDI 2024).

“Power limits and stricter regulation are making it harder to build new data centers. Meanwhile, GPU utilization is often below 50%. It’s time to get more from the GPUs already installed, and to put every job on the most cost-effective, efficient GPU that meets its demand,” said Alex Carter, Ph.D., Co-Founder & CEO, XPerf Inc.

ClusterReady: validate, proven with partners.
ClusterReady validates AI racks and clusters under AI training and inference workloads before they go live. A validation run takes 65 minutes to 8 hours, typically 3–4 hours, compared with 2–6 weeks for conventional L11/L12 testing. Each rack gets a tokens-per-second performance baseline, and every issue found comes with its root cause and remediation suggestions.

• With MiTAC Computing, ClusterReady showed an AMD Instinct™ MI325X rack performing 20% above AMD’s published FLUX.1-dev benchmark results in the tested configuration, with 95.8% scaling efficiency from one to four nodes on Llama 3.1 8B pre-training.
• With Akash Systems, ClusterReady documented the energy and thermal effect of Diamond Cooling^® on AMD Instinct MI350X servers under AI workloads.
• With AMD, XPerf operated MLPerf^® Training v6.0 submissions on AMD Instinct platforms in May 2026, on behalf of Akash Systems and MiTAC Computing.
• With a tier-1 ODM, XPerf finds the root cause of defective racks in minutes.

“XPerf gave us AI-workload performance evidence that demonstrates what our Diamond Cooling^® technology does in practice. We’re glad to keep working with XPerf because of the speed and quality of their work,” said Pamit Surana, Co-Founder & CCO, Akash Systems.

ClusterBeacon: operate, now available for commercial deployment.
ClusterBeacon learns each node’s safe operating envelope under load, then forecasts GPU cluster failures 10–30 minutes ahead, with recommended preventive actions. When something fails, it traces the incident to an evidence-cited root cause in minutes, where manual investigation typically takes hours to days. Engineers approve most changes, and ClusterBeacon verifies the fix with telemetry. It runs on-premises on NVIDIA and AMD GPU clusters, including in air-gapped data centers, and keeps all data on site.

ClusterMarshal: maximize, ready for pilot deployment
ClusterMarshal decides where, when and how each GPU job runs. It starts with operator input and moves toward fully autonomous placement.
Today a user describes a job: a training script or an inference model with a request rate and latency target. ClusterMarshal proposes two to four ways to run it, each with GPUs, start time, run time and cost. The user chooses one, and ClusterMarshal submits it as that user to the cluster’s existing scheduler. If it does not respond in time, the scheduler places the job as usual.
Pilots begin in shadow mode: ClusterMarshal recommends alongside the scheduler without affecting placement, so operators can compare its predicted run times and costs with what actually ran. Once operators trust the results, they enable submission for a queue or an entire cluster.
From there, ClusterMarshal makes the placement selection itself, submitting jobs without waiting for a person. Operators set the policies and limits, and the scheduler continues to execute each placement.
ClusterMarshal launches with Kubernetes support, with Slurm to follow. It runs on an XPerf-supplied appliance or customer-supplied hardware, including air-gapped environments; cloud deployment is also available. XPerf will demonstrate ClusterMarshal at SC26.

Availability:
ClusterReady and ClusterBeacon are available for commercial deployment. ClusterMarshal is available for pilot deployments, starting in shadow mode. Meet the XPerf team at SC26, booth 5617, November 17–19. To book a meeting, contact Alex Carter at alex.carter@xperf.ai.

About XPerf Inc.
Founded by ex-Intel engineers with deep experience deploying and managing GPU and ASIC clusters at scale, XPerf Inc. is an AI infrastructure software company building the control plane for heterogeneous data centers. Its products cover the full AI cluster lifecycle: the ClusterReady™ platform validates AI racks and clusters 10× faster than L11/L12 workflows; ClusterBeacon™ software predicts, prevents and resolves GPU cluster failures on-premises; and the ClusterMarshal™ scheduler helps teams choose where, when and how each GPU job runs. XPerf Inc. is a member of NVIDIA Inception, CNCF, the gBETA accelerator program, and the Endless Frontier Labs accelerator program. Learn more: xperf.ai

About Akash Systems.
Akash Systems is the pioneer in Diamond Cooling^® technology for AI servers and other heat-intensive electronic environments. First developed and deployed in partnership with NASA for satellites, Akash Systems is now providing its proprietary Diamond Cooling solutions for GPU, CPU, and memory systems to help address energy and performance challenges in the rapidly growing AI data center sector in the U.S. and globally. Akash is backed by leading venture capital investors Khosla Ventures and Founders Fund. Learn more at www.akashsystems.com.

XPerf, ClusterReady, ClusterBeacon and ClusterMarshal are trademarks of XPerf Inc. AMD, the AMD Arrow logo, AMD Instinct, and combinations thereof are trademarks of Advanced Micro Devices, Inc. Other names are for informational purposes only and may be trademarks of their respective owners.

Forward-Looking Statements: Statements about XPerf’s products, roadmap and pilot deployments are subject to risks and uncertainties, and actual results may differ materially. Performance results are from testing with the partners named, on the configurations stated, and do not guarantee results on other systems.

Lucia Dominguez Cabanes
XPerf Inc.
+1 607-438-8136
press@xperf.ai
Visit us on social media:
LinkedIn
YouTube

Media gallery