Compute Engine instances provisioning models

• Based on resource availability, you can immediately create compute instances.
• You can control when to stop or delete compute instances.

• Based on resource availability, you can immediately create compute instances.
• You can control when to stop or delete compute instances. However, you also allow Compute Engine to stop or delete compute instances at any time to reclaim capacity.

• Based on resource availability, you can create compute instances within a specified waiting time. For a standalone Flex-start VM, you can specify a waiting time of up to two hours. For a MIG resize request, the waiting time is indefinite.
• You can control when to stop or delete compute instances. However, you can't suspend or recreate them. Compute instances run for a minimum of 10 minutes and up to a maximum of seven days. When the compute instances reach the end of their run duration, Compute Engine stops or deletes them based on their termination action.

• You can request to reserve capacity at a future date for creating compute instances. If Google Cloud approves your request, then Compute Engine creates a reservation. At the start of the reservation period, you can consume the reservation by creating compute instances that match the reservation.
• During the approved reservation period, you can stop, restart, delete, and recreate compute instances to consume the reservation as needed. When the reservation period ends, Compute Engine deletes the reservation, and stops or deletes any compute instances that consume the reservation based on their termination action.

Ideal for workloads that require stability and continuous operation, such as the following workloads:

• Web servers
• Databases
• Enterprise applications
• Development and testing

Ideal for workloads that can tolerate interruptions, such as the following workloads:

• Batch processing
• High performance computing (HPC)
• Continuous integration and continuous deployment (CI/CD)
• Data analytics
• Media encoding
• Online inference

Workloads that require stability and need to run for no more than seven days, such as the following workloads:

• Small model pre-training
• Model fine-tuning
• HPC simulation
• Batch inference

Ideal for workloads that require stability and a specific run time, such as the following:

• For workloads that last up to 90 days:

  • Model pre-training jobs
  • Model fine-tuning jobs
  • HPC simulation workloads
  • Short-term expected increases in inference workloads

• For workloads longer than 90 days:

  • Training workloads
  • Inference workloads

Resource allocation varies based on how you create compute instances:

• Best-effort. If you create standalone Flex-start VMs, then Compute Engine physically places resources close to each other on a best-effort basis. To control placement, you can optionally use placement policies.
• Dense. If you create Flex-start VMs through MIG resize requests, then Compute Engine physically places resources on tightly coupled hosts connected by a high-speed network fabric to minimize network latency.

• You incur standard pricing for compute instances. See VM instance pricing.

• You incur charges based on the method that you use to create compute instances:

  • If you immediately create compute instances, then you pay as you go (PAYG).
  • If you create compute instances by using an on-demand reservation or an auto-created reservation for a future reservation, then you're charged until the reservation exists. For more information, see reservations billing.

• You get discounts up to 91% off for many machine types, GPUs, TPUs, and Local SSD disks. For more information, see Spot VMs pricing.
• You PAYG.

• Based on the machine series that your compute instances use, you get a discount as follows:

  • For A4, A3, and A2 machine series, you get a 53% discount for vCPUs, memory, and GPUs.
  • For H4D machine series, you get a 25% discount for vCPUs and memory.
  Other supported machine series aren't eligible for discounts. For more information, see Dynamic Workload Scheduler (DWS) pricing.
• You PAYG.

• You incur charges based on how you reserve capacity for creating compute instances as follows:

  • If you reserve capacity in AI Hypercomputer, then you incur charges based on accelerator-optimized VMs pricing. If you reserve resources for a year or longer, then you must purchase and attach a resource-based commitment to your reserved resources.
  • If you reserve capacity by using future reservations in calendar mode, then you incur charges based on the Dynamic Workload Scheduler (DWS) pricing.
• You're charged for the reservation period. For more information, see reservations billing.

To create compute instances, you must first reserve capacity using one of the following methods:

• To reserve capacity for long-running workloads, use future reservations in AI Hypercomputer.
• To reserve capacity for workloads that run for up to 90 days, use future reservations in calendar mode.

At your chosen delivery date and time, Compute Engine provisions your requested capacity. Then, you can consume the capacity by creating compute instances.

You can only use the following machine series:

• A4, A3, A2, G4, and G2 machine series
• N1 virtual machine (VM) instances with GPUs attached
• H4D machine series

Based on how you reserve capacity to create VMs, you can only use the following machine series:

• If you reserve capacity in AI Hypercomputer, then you can only use A4X Max, A4X, A4, A3 Ultra, A3 Mega, A3 High with 8 GPUs, A3 Edge, and H4D machine series.
• If you create a future reservation in calendar mode, then you can only use A4, A3 Ultra, A3 Mega, A3 High with 8 GPUs, or H4D machine series.

You can create compute instances as follows:

• Create a standalone compute instance.
• Create a MIG resize request.

Compute Engine uses DWS to schedule the provisioning of your requested capacity based on resource availability. DWS helps you obtain high-demand resources like GPUs.

• If you immediately create compute instances, then Compute Engine makes best-effort attempts to provision your requested capacity.
• If you create compute instances by consuming an on-demand reservation or an auto-created reservation for a future reservation, then you have very high assurance that Compute Engine provisions your requested capacity if the reservation has reserved capacity available.

You can control when to stop or delete a compute instance, except in the following cases:

• Compute Engine stops or deletes the compute instance to reclaim capacity. This process is called preemption.
• If the machine type that the compute instance uses doesn't support live migration, then Compute Engine stops the compute instance during host maintenance events.

Before a compute instance reaches the end of its run duration, you can do the following:

• Stop the compute instance: you can stop the compute instance at any time only if it was created as a standalone compute instance.
• Delete the compute instance: you can delete the compute instance at any time.

When a compute instance reaches the end of its run duration, Compute Engine deletes it.

You can control when to stop or delete a compute instance, except in the following cases:

• Compute Engine stops the compute instance during host maintenance events.
• The automatically created reservation to provision your requested capacity reaches the end of its committed reservation period. At that time, Compute Engine deletes the reservation, and stops or deletes any compute instances that consume the reservation based on the termination action that is specified in their configuration.