# Google Cloud previews smaller GPU slices in G4 virtual machines > Fractional G4 virtual machines introduce smaller NVIDIA GPU allocations, giving graphics and compute teams another way to size accelerator capacity. Canonical URL: https://www.devobs.io/news/news-google-cloud-fractional-g4-vms-preview/ By: Elias Brooks Published: 2026-09-06T11:58:54.637Z Updated: 2026-09-06T11:58:54.637Z Event date: 2026-03-17 Section: Infrastructure Google Cloud previewed fractional G4 virtual machines on March 17 as part of its NVIDIA GTC infrastructure announcements. The [announcement](https://cloud.google.com/blog/products/compute/google-cloud-ai-infrastructure-at-nvidia-gtc-2026) describes configurations that divide NVIDIA RTX PRO 6000 Blackwell Server Edition GPU capacity using NVIDIA virtual GPU technology. ## Smaller allocations for accelerator workloads The preview adds half, quarter, and eighth GPU options. Google positions these sizes across workloads including rendering, video processing, simulation, and remote desktops. It also describes managing fractional G4 machines with Google Kubernetes Engine and using Dynamic Workload Scheduler fallback priorities to find available slice configurations. The practical change is allocation granularity. A team whose application cannot use a whole GPU now has a smaller resource unit to evaluate. That does not establish that a fractional configuration will meet a particular latency target or deliver a proportional bill reduction; those depend on the workload and deployment choices. ## Measure the whole job For a graphics service, an evaluation should include concurrent sessions, frame delivery, startup time, and memory demand. For batch processing, measure completed work and queue delay as well as device utilization. A busier GPU is useful only if the service still finishes its work within the required window. Fallback capacity also needs an application policy. Decide which alternate sizes are acceptable and what should happen when none is available. Otherwise, an infrastructure mechanism for obtaining capacity can create unpredictable completion times for users. Treat the preview as an opportunity to characterize suitable jobs before committing a production dependency. Compare a representative workload on a full allocation and a smaller slice, retaining identical inputs and output checks. The resulting measurements can show where finer sizing removes unused capacity and where a larger allocation remains necessary. ## Source references -