Choosing between an RTX-class deployment and a data-center GPU such as an A100 is not a universal performance contest. Software requirements, memory, job size and interruption tolerance can change the answer. A small representative test is more useful than an unqualified savings claim.
1. Describe a successful run
Define the input, the output quality you need and the acceptable completion time. For inference, measure your own requests. For rendering, use a typical scene and its actual plugins. For processing, use a sample that resembles a normal day's files.
Check the software's GPU and memory requirements before choosing a machine. Include model weights, working memory and workload size. Adding replicas does not automatically turn several separate GPUs into one larger shared-memory machine.
2. Choose the service model
A container service runs your packaged application. A virtual machine or dedicated server provides a different operating environment. Confirm which one your software needs, who configures it and where data is stored.
For our managed application offer, we assess SaladCloud for suitable container workloads and help configure the customer's own provider account. Normandin Systems charges for the agreed setup or support; the customer pays the provider directly. See how our managed workload service works.
3. Test recovery as well as speed
Work that can retry in small batches is easier to recover than a single long job with no saved progress. Include a stop-and-restart test in your pilot and confirm that completed files are retained in the storage you selected.
Salad's availability figures are estimates, not a reservation for your job. Priority and workload requirements affect allocation. Check the provider's availability documentation when planning a launch.
4. Compare the full cost
Cost per usable result = total running, storage, transfer and service costs ÷ successful results.
Include all replicas, any retries, persistent storage, software licences and deployment help. Compare the same input, output quality and completion target on each option. Keep setup costs separate from recurring operating costs so you can see what changes with volume.
Salad's published compute rates are in USD and depend on configuration and priority. Check current provider pricing and the actual billing configuration before approving a test. A quoted CAD service fee and a USD compute bill need to be compared in a common currency.
A short checklist before paying
- The application and required licences work in the proposed environment.
- GPU memory and other resources fit a representative sample.
- Files, progress and outputs have an agreed storage location.
- Recovery is tested and ongoing operating responsibilities are assigned.
- The pilot has a budget, a stop procedure and a definition of success.
You do not need a large initial deployment to answer these questions. Start with a bounded pilot, measure it and then decide whether scaling makes sense.
Provider information checked October 6, 2026. Prices and capacity can change. This guide does not quote a hardware allocation or guarantee a saving.