Flavors
An automated system checks the flavor catalog on this page against live Hyperstack data every day. When the live data changes, the documentation team updates the page. Learn how these docs stay accurate.
A preset combination of GPU, CPU, RAM, and disk that you select as a single unit when sizing a workload.
A flavor is a preset combination of GPU, CPU, RAM, and system disk that you select as a single unit. Instead of configuring each component on its own, you choose a flavor and Hyperstack provisions the exact hardware it specifies. The flavor you choose determines the compute, memory, and storage your instance receives, so it is the first decision when you size a workload.
You select a flavor whenever you deploy a virtual machine or provision a Kubernetes cluster. In a Kubernetes cluster, worker nodes accept any flavor, while master nodes require a CPU-only flavor.
What a flavor includes
Every flavor defines a fixed combination of the following:
- GPU model and count: the GPU type (for example, a Blackwell B200 or a Hopper H100) and how many are attached to the instance. CPU-only flavors include no GPU.
- CPUs: the CPU cores allocated to the instance.
- RAM: the system memory available to the instance.
- Root disk: persistent local NVMe storage that holds the operating system and essential files for the life of the VM.
- Ephemeral disk: ephemeral storage, a temporary drive attached for the active workload. Ephemeral data is non-persistent and is wiped when the VM is hibernated or deleted.
- Region: the data center where the flavor is available.
Each flavor also supports a specific set of capabilities, known as flavor features, that determine how you can operate the instance. These include network-optimized bandwidth, hibernation, snapshots, and boot from volume.
Flavor configurations by GPU model
The catalog below lists every flavor Hyperstack offers, grouped by GPU family. Each entry shows the GPU count, CPU cores, RAM, root disk, ephemeral storage, region, and supported features for that configuration.
Filter by family, region, GPU count, or supported feature, or search by name.
To check whether a flavor's GPUs are in stock before you deploy, see live GPU stock or the Deploy Virtual Machine page in the Hyperstack console, which shows live stock and pricing together.
Blackwell-generation SXM for the largest training and inference workloads.
B300-SXM#
n3-B300-SXM6x8Blackwell-generation SXM with 192 GB HBM3e for the largest training and inference workloads.
141 GB HBM3e and 4.8 TB/s bandwidth for large-model training and inference.
80 GB Hopper for AI training, inference, and HPC across PCIe and SXM variants.
80 GB Ampere for AI training, data analytics, and HPC in PCIe and SXM configurations.
RTX PRO 6000
Blackwell server GPU for AI inference, rendering, and virtual workstations.
Ada Lovelace GPU for visual compute, real-time ray tracing, and AI inference.
RTX A6000
Ampere professional GPU for AI, rendering, simulation, and multi-GPU workloads.
RTX A4000
Cost-efficient Ampere for deep learning and real-time ray tracing.
GPU-less compute for web services, networking, and databases.
Networking
Every flavor includes virtual NICs with up to 16 Gbps of inter-VM bandwidth by default.
Flavors marked 350 Gbps in the catalog can be deployed into network-optimized environments that raise inter-VM bandwidth to 350 Gbps via SR-IOV. Network-optimized environments are required for distributed training across multiple VMs and for any workload that saturates the default vNIC.
N3 generation
The N3 series is the current generation of Hyperstack flavors. Compared to N1 and N2, N3 flavors offer:
- Shorter provisioning times for new VMs, especially flavors with large root disks (such as those used for Kubernetes nodes).
- Faster deletion of VMs.
- 10 to 15 percent higher CPU and memory performance.
- Inter-VM bandwidth up to 350 Gbps on select GPU flavors when deployed into network-optimized environments. See high-speed networking.
Large root disk flavors
If you need more persistent disk than the flavors above provide, Hyperstack offers Large Root Disk Flavors that you can deploy through the API. These flavors reallocate the ephemeral disk capacity into the root disk, giving you a single, larger persistent disk that lasts for the life of the VM. They suit high-I/O, containerized, and short-lived workloads. Because they have no ephemeral disk, they do not support snapshotting or hibernation.
Related
Features and restrictions
Hibernation, snapshots, boot-from-volume, and hard-reboot support per flavor.
Large root disk flavors
API-only variants with larger persistent root disks in place of ephemeral storage.
GPU stock information
Live GPU availability per region and the stock reporting API.
Regions
The data center locations that house each flavor.
Spot virtual machines
Spot variant features, trade-offs, and best-fit workloads.
High-speed networking
Network-optimized environments that raise inter-VM bandwidth to 350 Gbps via SR-IOV.