GPU types and blocks

GPU types. Hangar offers four GPUs in Australia, and this page shows the command that starts each one. You pay from prepaid credit at the price of the GPU, including GST: A$1.35 an hour for a fractional L4 and A$4.55 an hour for an L4, by the second. An H100 or a Trainium2 costs the price that hangar status shows, by the minute.

Start now on an H100 or a Trainium2: by the minute, no booking

A block is a set of GPU capacity that Hangar reserves. It fixes the GPU type and the GPU count of every session in the block. Hangar's H100 blocks in Sydney and its Trainium2 block in Melbourne start when you ask, while a machine of the block is free. Each session is a fresh instance that Hangar launches in the block for your organisation alone. Hangar attests and resets the instance before your job starts. The session ends with a job certificate.

hangar status --pools
hangar run --gpu trn2 model:v1 -- python infer.py
hangar run --gpu h100 --limit 60 train:v2 -- python train.py

• What is free now. hangar status shows each pool, and hangar status --pools shows the pools without a sign-in. For each pool, the output shows whether the pool is free now and until when, or when the pool is free again. The output also shows the price including GST. The CPU sandbox shows its price for each vCPU-hour.
• Price. The price is the price of the pool for each accelerator hour, including GST. hangar status and the pricing page show the price. Hangar takes credit for each minute that starts, from the moment the instance runs to the moment Hangar terminates the instance.
• Length. A session ends at the first of these events: its jobs end, your --limit passes, you run hangar stop, your credit reaches its last three minutes, or the switchover before the next booking on its machine begins. When the session starts, Hangar shows the time the session ends.
• A block of more than one machine. A block holds one to four machines of the same instance type. Each machine takes one session at a time. A block of two machines takes two sessions at once, each on a fresh instance of its own. Hangar starts your session on the first machine that is free. Before the launch, Hangar asks the provider for a free instance in the block's reservation. Before your job starts on a machine, the wipe record of the previous instance on that machine must exist.
• When the block is busy. Every machine of the block is busy. Hangar starts nothing and charges nothing. Hangar shows the earliest time a machine is free again. You can start a job then, or book ahead with hangar book.
• Bookings. A booking keeps its time. A session never runs into a booking, or into the 15-minute switchover before the booking.

Instant GPUs: by the second, no booking

Two GPUs start when you ask. You do not book a night. Each session has its own instance for your organisation. The worker attests the instance, and Hangar resets the instance before your job starts. The session ends with a job certificate. The certificate shows the instance type and the GPU.

An L4 starts on g6.xlarge in the first Sydney zone that has capacity. When no Sydney zone can start g6.xlarge, Hangar starts g6.2xlarge. The g6.2xlarge has one NVIDIA L4 with 24 GB of GPU memory, the same as the g6.xlarge, and you pay the same price for each hour. The certificate records the instance type that started.

The fractional L4 is one eighth of a physical NVIDIA L4. Each fractional session has its own kernel, driver, memory, disk and network interface. Other instances share the physical GPU through vGPU (NVIDIA's way to split one GPU), and the provider keeps the shares apart. When the session starts, Hangar sets the GPU memory of your share to zero and reads the memory back. Choose the L4 when your work needs a GPU that no other instance shares.

hangar run --gpu l4-fraction smoke:v1 -- nvidia-smi -L
hangar run --gpu l4 --limit 90 server:v3 -- python serve.py
hangar stop <session>

• Metering. Hangar meters by the second, from the moment the instance runs to the moment Hangar terminates the instance. The certificate records both times. Hangar takes credit one minute at a time, as each minute starts. Hangar returns the unused part of the last minute when the session ends.
• Starting. A start needs credit for ten minutes at the price of the GPU. The new-account limits apply to a start as they apply to a booking. The session runs the command in the start. You cannot add a later job to the session.
• No job, no charge. If the instance fails to attest, to reset or to stage, no job starts and Hangar does not charge the session.
• Ending. A session ends at the first of these events: its jobs end, the limit you set with --limit (in minutes from the start) passes, you run hangar stop, or your credit reaches its last three minutes. Ten minutes before Hangar stops a session for credit, Hangar shows a warning. When a session stops, your job receives SIGTERM and has 60 seconds to save its work.
• Capacity. Hangar starts the session in whichever Sydney zone has the GPU at that moment. When no zone has the GPU, Hangar starts nothing and charges nothing. Hangar then shows the next night that has a block, and you can book that night.
• Use. The fractional L4 suits GPU tests in continuous integration and CUDA smoke tests. The L4 suits inference with models of up to about 8 billion parameters in 16-bit precision, and small LoRA fine-tuning.

hangar status, hangar logs, hangar pull and hangar verify accept a session id. They accept a booking id too.

Read this page as Markdown