By news editor
Google Cloud has made Cloud Pathways generally available on Google Kubernetes Engine, its managed Kubernetes service. The move puts in customers’ hands the distributed machine-learning runtime that Google DeepMind built to train and serve Google’s own models, Gemini among them. A runtime here is the software layer that coordinates one training or inference job across many AI chips at once.
In the blog post announcing the release, Google says the software already carries heavy use. Over the last 90 days, by its count, AI companies and enterprises processed more than 130 million Cloud TPU chip-hours using Pathways on GKE. TPUs are Google’s custom AI accelerator chips.
What changes for engineering teams is how a large job is controlled. In the default setup for JAX, the Google-developed machine-learning framework that Pathways supports, every TPU host runs its own Python interpreter and sees only the devices attached to it. Pathways replaces that with one controller. A single client process, the post says, can orchestrate multiple large groups of chips “spanning thousands of TPU chips as if they were all connected locally.”
The business case Google makes is about wasted accelerator time. Elastic training lets a job keep going when hardware fails: the runtime uses snapshots held in host memory to shrink the job onto the remaining chip groups without crashing the client, according to the post. A second feature, the Shared Pathways Service, lets several users and jobs draw on one long-running pool of TPU capacity. Helper containers running on the TPU hosts take on Python work such as data loading and checkpointing.
Customer evidence in the announcement is thin on names. Google identifies one customer, ZenteiQ, which it says uses Cloud Pathways on GKE to pre-train scientific foundation models and run reinforcement learning across clusters of Trillium TPUs. The post carries a statement from ZenteiQ founder Sashikumaar Ganesan saying the company’s machine-learning engineers no longer spend their hours watching job health or managing manual restarts. That statement includes no recovery times or cost figures.
The largest user Google describes goes unnamed. The post says one of “the world’s largest consumer technology companies” runs production batch and reinforcement learning inference on Cloud Pathways across tens of thousands of TPU chips from three generations, consuming over 100 million chip-hours per quarter. That company also co-designed elasticity features with Google as it expands into pre-training, the post says.
Set side by side, the two numbers suggest concentration. Google cites more than 130 million chip-hours across all customers in 90 days and over 100 million per quarter for a single customer. The post does not say whether the two figures cover the same period or the same chips, so the share held by that one customer cannot be calculated from it.
Several things a buyer would want are absent. The announcement gives no price for Cloud Pathways and does not say whether the runtime carries a charge beyond the TPU capacity it runs on. It offers no benchmark for how much idle time elastic training saves. It describes JAX workloads throughout and makes no mention of PyTorch, another machine-learning framework, or of GPUs. The customers positioned to gain are those already running on Google’s TPUs.
Availability is broad within that limit. Google says Cloud Pathways is generally available in all Google Cloud regions that support four TPU types on GKE, namely TPU v5e, TPU v5p, Trillium and Ironwood, with container images that support JAX 0.10.1 and later. An integration with Cluster Director, which extends Pathways to clusters managed with the Slurm scheduler, is in Preview, Google’s label for pre-release features.
The next marker is the Slurm integration reaching general availability, which would open the runtime to teams that schedule their work with Slurm instead of Kubernetes. Pricing terms and a second named customer would show whether usage is spreading beyond the one large account Google describes.
Leave a Reply