What a slot gets
A slot gives you one GPU instance for the booked window, and nobody else uses it. You pay for the booked hours, including GST, from prepaid credit.
The block sets the GPU type and the GPU count. Every slot line prints them. A job always uses the hardware it was booked on.
The instance
| Instance type | GPUs | vCPU | Memory | Local NVMe | Network |
|---|---|---|---|---|---|
| p5.4xlarge | 1 × NVIDIA H100, 80 GB HBM3 | 16 | 256 GiB | 1 × 3.84 TB | 100 Gbps EFA, without GPUDirect RDMA [verify] |
| p5.48xlarge | 8 × NVIDIA H100, 80 GB HBM3 each, NVLink | 192 | 2,048 GiB | 8 × 3.84 TB | 3,200 Gbps EFA, with GPUDirect RDMA [verify] |
On p5.4xlarge, NCCL over EFA needs FI_HMEM_DISABLE_P2P=1 in the job's environment.
What each job gets inside the slot
| Resource | p5.4xlarge | p5.48xlarge |
|---|---|---|
Memory limit (memory.max) | 240 GiB: the instance's memory less 16 GiB for the host | 2,032 GiB |
/dev/shm | 64 GiB, a quarter of memory | 512 GiB |
| Swap | Off | Off |
| Locked memory | Unlimited | Unlimited |
| Transparent huge pages | On request (madvise) | On request (madvise) |
| Image size | Up to 200 GB compressed | Up to 200 GB compressed |
| Working disk | The local NVMe, encrypted with a key held only in memory | All eight NVMe devices as one encrypted volume |
Paths inside the container
/hangar/outputs: Hangar seals everything written here to Sydney storage when the job ends, and lists it on the certificate./hangar/checkpoints: all jobs in the slot share this folder. Hangar seals it after each job and when the slot ends, and keeps it for your next slot./data/...: the uploads that the job lists. Hangar checks their digests, then mounts them read-only.- Each secret arrives under its name, as an environment variable or as a read-only file.
Time
The job receives SIGTERM 60 seconds before the slot ends, and SIGKILL at the end. If the job saves to /hangar/checkpoints on SIGTERM, it resumes from there in your next slot.
If a job reaches a memory or disk limit, the job ends abnormally. The certificate gives the reason, and Hangar keeps the last checkpoint.
Network
A job reaches the names on the account's egress allowlist over HTTPS (port 443). Hangar stages the job's data from Sydney storage before the job starts.
Inside the slot, the worker looks up each listed name and passes the job's connection to that name unchanged. The connection stays encrypted from the job to the service.
Hangar refuses each name lookup or connection outside the list, and records the refusal on the job certificate. Allowlist entries are host names, such as huggingface.co. Set them in the account policy.