Instance types & specifications

Mithril offers several instances of various NVIDIA GPU types fit for your workloads. We will continue to add instance types over time.

Depending on your instance type, Ethernet or InfiniBand interconnect may be available.

VM image

You can choose between one of the supported VM images. We recommend Ubuntu24.04 2025.10.10 for the latest software, best performance, and updated security. This image is the default if none is specified in the launch specification.

Image name	NVIDIA driver	CUDA	Notes	NVIDIA container toolkit	OS
Ubuntu24.04 2025.10.10	580.95.05	13.0	Improved fileshare performance	1.17.8	Ubuntu 24.04 LTS
Ubuntu24.04 2025.08.01	570.86.10	12.8	Does not support executing binaries from persistent storage	1.17.8	Ubuntu 24.04 LTS
Ubuntu22.04 2024.07.24	535.183.01	12.2			Ubuntu 22.04

You can use startup scripts to customize your environment during the startup process of your instances.

Hardware specifications

Each instance comes with its own ephemeral storage (outlined below) with NVMe SSD. This storage will exist until the instance has been terminated or pre-empted. There is no cost to use the ephemeral storage. Ephemeral storage is automatically mounted for you (and re-mounted after preemptions) at /mnt/local.

Achieve 25% faster start-up times when configuring your instances with RAM lower than the default! RAM can be specified in the console or API, with smaller amounts enabling faster start-up time. Note the compute price does not change based on how much RAM you have — pricing is per GPU hour.

Instance	Region	GPU type	GPUs	VRAM per GPU (GB)	CPU cores	System RAM (GB)	Boot disk (GB)	Ephemeral NVMe (GB)
8x H100 80GB SXM5	us-central2-a	H100 SXM	8	80	168	1,814	256	12,500
1x H100 80GB SXM5	us-central2-a	H100 SXM	1	80	20	226	256	1,774
8x H200 141GB SXM5	us-central2-b	H200 SXM	8	141	168	1,600	256	12,500
1x H200 141GB SXM5	us-central2-b	H200 SXM	1	141	20	200	256	1,500
8x H200 141GB SXM5	af-south1-a	H200 SXM	8	141	160	1,825	256	15,500
1x H200 141GB SXM%	af-south1-a	H200 SXM	1	141	20	227	256	3,300

Memory (RAM) configuration

To improve Spot startup times by up to 25%, we recommend launching bids with less RAM configured in the launch specification. This is ideal for light workloads or development use cases that require fast startup but low resource utilization.

Use the console or API to launch instances with less RAM.

At any point after launch, RAM can be adjusted from 100GB up to the instance type's default. In order to adjust the RAM for an order, the bid must be paused and resumed, which will restart all instances in the order.

By default, instances are launched with the maximum memory (RAM) allowed for the instance type.

Note that this feature is only available for Spot instances at this time.