Together GPU Clusters
A GPU cluster service offered by Together AI that supports large-scale training workloads and provides operational capabilities such as automatic node repair and fine-grained control by fault type and Xid code.
Timeline (last 90 days)
Together GPU Clusters has launched an auto-remediation feature, aiming to return faulty nodes to a healthy state automatically as much as possible while leaving remediation decisions to the user.
The auto-remediation system covers three types of fault sources: the GPU itself, the host, and the cluster fabric.
After detecting a problem, the system chooses "the lightest fix," escalating remediation measures step by step according to fault severity rather than always applying the heaviest remedy.
Users can individually decide whether to enable auto-remediation for each type of fault and each Xid error code.
The information Together disclosed this time focuses on the mechanism itself, and the specific list of remediation actions and the Xid coverage still await more detailed documentation.
Together GPU Clusters has announced support for preemptible compute, offering equivalent GPU resources at a fixed 50% discount off on-demand pricing.
All Facts (6)
The auto-remediation system covers three types of fault sources: the GPU itself, the host, and the cluster fabric.
50%UnverifiedAfter detecting a problem, the system chooses "the lightest fix," escalating remediation measures step by step according to fault severity rather than always applying the heaviest remedy.
50%UnverifiedUsers can individually decide whether to enable auto-remediation for each type of fault and each Xid error code.
50%UnverifiedThe information Together disclosed this time focuses on the mechanism itself, and the specific list of remediation actions and the Xid coverage still await more detailed documentation.
50%UnverifiedTogether GPU Clusters has launched an auto-remediation feature, aiming to return faulty nodes to a healthy state automatically as much as possible while leaving remediation decisions to the user.
50%UnverifiedTogether GPU Clusters has announced support for preemptible compute, offering equivalent GPU resources at a fixed 50% discount off on-demand pricing.
50%