3 ms·
GPUs in the cloud right now are so pricy, in part I think because NVIDIA is making people put in K40s and K80 - the Quadro and Telsa class cards that are super
by bhouston 10y ago
GPUs in the cloud right now are so pricy, in part I think because NVIDIA is making people put in K40s and K80 - the Quadro and Telsa class cards that are super expensive.
We started to put on premise machines that have GPUs in them to reduce our costs. We can get a desktop class Core i7 machine with a GeForce 1060 6GB for around $800 all in, where as on Amazon you have to have a G2 2.xlarge on demand it is $400/month, or $3478 one year term (that is for 1/4 of a G2 8.xlarge achine, thus you get one of its 4 Quadro K5000 4GB cards to yourself.) I could put 2 1060s in that machine for another $200 or so. This just beats AWS pricing in <4 months of use.
My expectation is that these 1060s will be good for two years, maybe even three.
Because our GPU tasks are low bandwidth and non-critical we can run them on our existing office internet. Our rent also includes power. So there are not too many additional costs over the bare machine costs.
- mabbo 10y agoI think you're definitely winning on price- provided you plan to run things continuously for that time period. The AWS GPU instances are probably not aimed at you though. It's far more likely to be aimed at the Master's student in ML who has this new awesome model but it takes goddamn forever to train on his laptop and he just needs 3 days of time with a big-ass GPU instance. He spins up the instances, runs his model, spins it down and is out what, $100 maybe? That fits into his limited budget. Additional benefit: if I buy a kickass GPU, I'm going to reduce my overall productivity in research because now I can finally play some of these hot new games on full graphics!
- tmsldd 10y agoHopefully everything works fine for the student at the first run ;)
- bhouston 10y ago> It's far more likely to be aimed at the Master's student in ML I am pretty sure AWS isn't targeting its services at Master's students.
- chrisfosterelli 10y ago> It's far more likely to be aimed at the Master's student in ML As a counter example, I'm a master's student in ML and my lab wants to push more for an onsite machine since it's easier to expense one machine than a bunch of monthly AWS bills for each student in the lab :) Additionally, most Master's students (at least in Canada) already have access to one or more very large compute clusters for jobs like this. I have used AWS a few times when I was short on time though!