Choosing between Forge 80 and Anvil 192

We get this question in support every week. The short answer: pick by memory first, then by price per finished job, not price per hour.

When Forge 80 is the right call

Forge 80 has 80 GB per GPU and costs $2.29 an hour. For fine-tuning models up to about 30 billion parameters, and for most runs that fit in memory with sharding, it finishes the job for the least money.

When Anvil 192 pays for itself

Anvil 192 has 192 GB per GPU at $4.49 an hour. It earns its price when a model or its context no longer fits on Forge without heavy sharding, because fewer GPUs means less time spent talking between them.

• Long-context training, where activations grow with sequence length

• Serving a large model on one GPU instead of splitting it over two

• Jobs that spend more than a third of each step on communication

Measure it on your own job

Run 50 steps on each and compare cost per step, not speed per step. This is the script our support engineers use:

from ferrite import Cluster

for gpu in ["forge-80", "anvil-192"]:
    c = Cluster.create(gpu=gpu, count=8, region="ams-1")
    r = c.benchmark("train.py", steps=50)
    print(gpu, r.seconds_per_step, r.cost_per_step)
    c.delete()

from ferrite import Cluster

for gpu in ["forge-80", "anvil-192"]:
    c = Cluster.create(gpu=gpu, count=8, region="ams-1")
    r = c.benchmark("train.py", steps=50)
    print(gpu, r.seconds_per_step, r.cost_per_step)
    c.delete()

from ferrite import Cluster

for gpu in ["forge-80", "anvil-192"]:
    c = Cluster.create(gpu=gpu, count=8, region="ams-1")
    r = c.benchmark("train.py", steps=50)
    print(gpu, r.seconds_per_step, r.cost_per_step)
    c.delete()

Price per hour is what you see on the invoice. Price per finished run is what you actually pay.