
We initially migrated some workflows to it due to cost. After actually modifying code to deploy on Salad infrastructure, we found many of its purported benefits were silently qualified or inaccurate. Review collected by and hosted on G2.com.
There is a long list here, but here are the highlights that users should definitely know:
After running 3 instances to run batch data-processing workflows at 50mbps traffic each for several weeks, I received an email from them saying that amount of bandwidth necessitates a $0.15/gb charge that they're going to also back-bill me for to the beginning of the month. We're talking ~$7000/mo in traffic charges here for running 3 instances that tap out around 50-60mbps. And when I tell them no, they tell me it is part of a larger plan to introduce traffic charges to users. As a contrast to that, I also run instances on Vast and get charged less than $2/mo for similar network traffic that runs in the 700-1200mbps range. There is nothing odd we are doing with this workflow. Its simply a heavy data-processing workflow, and frankly you can't even find commercial services that run at 50-60mbps because it is too slow to pay for. I can publish the emails sent about this if anyone wants further verification.
Instance performance - Instances are interruptable and don't perform anywhere near the same throughput as instances on other providers. We tried deploying on multiple RTX5090 instances (that we also have deployed on other providers). There was clear attractiveness in the pricing, however after actually deploying them we found that they performed at around 20% of the max throughput of our non-Salad infrastructure. We have some custom workflows that batch process data with a quantized Qwen model with a low token cap. On 2 other providers we process 30k-33k requests/hr through the stack. On Salad 5090 instances, every one we tried capped out at around 5k-6k/hr. So, we would have to deploy ~6 5090 instances on Salad to reach the throughput of 1 on our other providers. During research to try to solve this, it largely came down to the instances operating in a WSL environment, with some also being because of slow bandwidth from the instances or inadequate power. Ironically, we're talking 15-60mbps bandwidth that they now want to charge ~$7k/mo to use across 3 instances.
That causes quite a shift in in the value proposition...
From: "we're cheaper and decentralized"
To: "we'll provide 10% of the bandwidth and 20% of the GPU throughput at slightly cheaper than competitors, unless you actually use 50mbps network traffic all month - in which case we'll say we're going to back bill you thousands of dollars at an absurd rate for network traffic that is literally too slow to buy. But, also be aware that the 5090 pricing will also necessitate you running 6 instances to get the same throughput as 1 on other providers, and you'll have to run more than 1 because they're all interruptable."
Inventory - The available instance inventory is highly inaccurate. Initially deploying on instances their inventory said had more than 50 available would stay un-allocated because those instances either weren't actually available or some other issue with their allocation engine.
Customer service - Prior to deploying I had sent some questions to the support inbox. The last support question I asked them, they responded to after 8-9 days.
The experience is accurate and speaks for itself. Review collected by and hosted on G2.com.
