Featherless GPU Cloud · AMD Instinct MI325X

Rent MI325X GPUs
by the hour

256GB of memory on every GPU. Launch 1, 2, 4 or 8 from the console or the CLI and pay per second while they run.

MI325X256 GB HBM
Memory per GPU
256 GB
Memory bandwidth
6.0 TB/s
GPUs per instance
1, 2, 4 or 8
Access
Root SSH, terminal, JupyterLab
CPU and RAM per GPU
8 vCPU, 256 GB
Beta price
$3.29 / GPU-hr
256GB
On one GPU

A job that needs three 80GB cards and a sharding config fits on one device.

$3.29
Per GPU-hour

$2.99 per GPU on a full node. Billed per second from prepaid credit. A stopped instance is not billed for compute.

Self-serve
Sign in, add credit, launch

The console shows what is available and holds it when you launch. The CLI launches the same instance from a script.

One node8 of 8 GPUs
GPU 1256 GB
GPU 2256 GB
GPU 3256 GB
GPU 4256 GB
GPU 5256 GB
GPU 6256 GB
GPU 7256 GB
GPU 8256 GB
Rent 1 GPUNode ceiling · 2,048 GB
Why MI325X

Size the job by memory.

Memory capacity decides whether a model fits. MI325X carries 256GB on a single card, so work that would be split across three 80GB GPUs runs on one, with no sharding config and no interconnect to tune.

A full node holds eight: 2,048GB of GPU memory, 64 vCPUs and 2,048GB of system RAM in one instance, with 6.0 TB/s of memory bandwidth per card.

What people rent it for

Hourly suits work that finishes.

A fine-tune runs for hours or days and then stops. Billing is per second and ends when the container does, so between jobs you pay for stored volumes and nothing else.

Fine-tuning

Adapt an open model to your own data. Start from the PyTorch (ROCm) template, write checkpoints to a network volume, and stop the instance when the run ends.

Training and pre-training

Take a full node: eight GPUs and 2,048GB in one instance. Checkpoints on a network volume survive a stop, a restart and a deleted instance.

Evaluating AMD

Run your own workload on MI325X before you commit to it. Serve it with the vLLM (ROCm) template behind a public HTTPS port. If it holds up, managed GPU is the same hardware.

Batch jobs and sweeps

Submit one command as a Run Job from the CLI. It streams output back and releases the GPU when the command exits. You pay only while the command runs.

Beta price
$3.29
per GPU-hour

$3.19 at two GPUs, $3.09 at four, $2.99 at eight. Metered per second from prepaid credit. The rate you launch at holds for that instance, and there is no contract.

Pricing

What each configuration costs.

Add credit by card and pay per second while the container runs. Waiting for capacity and pulling the image are free. A stopped instance is not billed for compute.

Configuration
GPU memory
Per GPU-hour
Per hour
1× MI325X
256 GB
$3.29
$3.29 / hr
2× MI325X
512 GB
$3.19
$6.38 / hr
4× MI325X
1,024 GB
$3.09
$12.36 / hr
8× MI325X, full node
2,048 GB
$2.99
$23.92 / hr

Beta prices as of 19 November 2026. Each GPU comes with 8 vCPUs and 256GB of system RAM. Container disk costs $0.30 per GiB a month while the instance runs. Network volumes cost $0.30 per GB a month until you delete them.

What the same card costs elsewhere

Provider
MI325X on demand
Terms
Featherless
$3.29, or $2.99 at 8 GPUs
Billed per second. No compute charge while stopped. 1, 2, 4 or 8 GPUs.
DigitalOcean
$3.80
Billed per second with a five-minute minimum. Billed while powered off. 1 or 8 GPUs.

DigitalOcean rate from its published pricing, checked 19 November 2026.

Memory per dollar

GPU
On-demand rate
Per GB of memory, per hour
MI325X on Featherless, 256 GB
$3.29
1.3 cents
H100 on Lambda, 80 GB
$4.59
5.7 cents

Three H100s hold 240GB for $13.77 an hour. One MI325X holds 256GB for $3.29. Lambda rate from its published pricing, checked 19 November 2026.

How it works

Signup to running job, without a call.

Step one

Sign in and add credit

Use your Featherless account or create one, then add credit by card. One hour of an instance's rate is enough to start it.

Step two

Pick a template and a size

PyTorch, vLLM and SGLang templates ship with ROCm, SSH and JupyterLab, or bring your own linux/amd64 image. Choose 1, 2, 4 or 8 GPUs and mount a network volume for anything you want to keep.

Step three

Connect and run

SSH in as root, open the browser terminal, or work in JupyterLab. Stop the instance when the job ends and compute billing ends with it.

Or from the CLI
featherless-cloud instance create --name quickstart --template featherless/rocm-pytorch --gpu-count 1
FAQ

Questions before you launch

Will PyTorch and my training stack run on AMD?

If nothing in it depends on CUDA-only code, generally yes. The PyTorch (ROCm) template ships Python 3.12, PyTorch 2.12 and the ROCm libraries. PyTorch's torch.cuda API drives the GPU through ROCm, so torch.cuda.is_available() returns True. A hard CUDA dependency is the one deal-breaker, because the port costs more than the rate saves. If you are unsure about a library, ask in Discord before you add credit.

How does MI325X compare to H100 and H200?

On memory, 256GB per GPU against 80GB on an H100 and 141GB on an H200. We have not measured throughput against either, so we do not quote it. Billing is per second: ten minutes of your own job on one GPU costs about 55 cents.

Can I train across multiple nodes?

No. One instance or Run Job uses up to eight GPUs on a single node, 2,048GB in total. You can run several side by side, up to 16 GPUs per organization by default, and ask in Discord for a higher limit.

Is there an SLA?

Not on GPU Cloud. That is fine for a training run and wrong for production traffic with a latency target. If you need an uptime commitment, managed GPU is the product, and it runs on the same hardware.

Do you manage anything for me?

We maintain three ROCm templates, PyTorch, vLLM and SGLang, each with SSH and JupyterLab. Everything after launch is yours: the model, the settings and the job. If you want a model catalog behind an API, that is our serverless inference product.

Can you take my GPU away mid-job?

Not from a running instance. It sits on dedicated compute until you stop it or your credit runs out. If you stop it, starting again needs a free GPU of the same model, which can fail when capacity is short. Run Jobs are different: they stop at 24 hours and can be interrupted to make room for other work, so allow retries and write checkpoints to a network volume.

Do you look at my data?

No. It is your data and we do not read it. We monitor metadata only, to catch abuse such as someone attacking a third party from our network. Files on a network volume stay on our disks until you delete the volume, so we have legal obligations and would cooperate with a lawful request.

How does billing work when I stop?

Usage is metered per second from the moment the container starts. Waiting for capacity and downloading the image are not billed. A stopped instance is not billed for compute, and its container disk bills only while it runs. Network volumes bill for their full size from creation until you delete them.

What happens to my files when I stop?

The container disk is erased on every stop and restart, installed packages included. Keep datasets, checkpoints and the Hugging Face cache on a network volume. A volume keeps its data when the instance stops or is deleted, can grow while mounted, and can be shared by several instances.

Is there a minimum spend?

There is no commitment. An instance starts once your credit covers one hour of its rate, which is $3.29 for one GPU. After that you pay per second.

What is not published yet?

The region and egress pricing. If either decides it for you, ask in Discord or email gpu@featherless.ai before you add credit.

Self-serve

Deploy your first GPU

Sign in with your Featherless account, add credit and launch. The quickstart covers sign-in, launch, running code and clean-up.

Sales-assisted

Renting at volume?

If you are moving an inference workload or need more than 16 GPUs, talk to us about managed GPU or a higher limit.

  • Paid evaluation on MI325X before you commit
  • Managed GPU once the workload settles into production
  • Sustained capacity on a term that suits the work