Utilising GPU's on Amazon AWS instead of CPU

robin
New Altair Community Member
In terms of the recommended set-up on AWS for RM it is the m4.xlarge with 4 CPU's and16 GB RAM. These are standard CPUs and I was wondering if anyone had used the G3, P2 or P3 instances available for having access to the GPU processing abilities on the platform?
Have you noticed any performance increase with GPU's? Is cost benefit analysis efficient in terms of the extra cost for the GPU vs the reduced model training time? How easy has it been to run in the cloud vs dedicated server?
Have you noticed any performance increase with GPU's? Is cost benefit analysis efficient in terms of the extra cost for the GPU vs the reduced model training time? How easy has it been to run in the cloud vs dedicated server?
1
Best Answers
-
I had good results with AWS. At some point my bills exceeded the cost to go dedicated so that was the reason for the switch.
2 -
nice conversation here. Yes when I was freelancing I had EC2 instances running RM Server for clients that I had on a Lambda CRON sched. It would spin up once/day, do its thing, and spin down. MUCH cheaper than leaving it on 24/7.
Has anyone done an apples-for-apples benchmarking analysis using a variety of EC2 instance types with a fairly typical ML modeling task (not DL)? I sometimes feel like I'm just grabbing some config at random sometimes..
Scott
1
Answers
-
Only the Keras and DL4J extensions can make use of GPU's. Unless you are using those extensively there is no benefit to those instances. Even then there are some good reasons to stick to the CPU as the GPU can still be memory limited. I have done some testing of GPU's against high core count CPU's and the advantages are not always there. I have also compared AWS to dedicated and ended up going with dedicated but dedicated makes sense if you need to run applications 24/7. If it is just for Rapidminer or general ML tasks then AWS works very well.3
-
We go through spurts of running processes, with the server standing idle for days between the executions. In some sense it makes more sense to spin up an instance and use it when required and then turn it down again when everything is complete. We do run applications 24/7 for a couple of days during the month, as well as cron jobs but that would be handled by the EC2 instance while it is live.1
-
I had good results with AWS. At some point my bills exceeded the cost to go dedicated so that was the reason for the switch.
2 -
nice conversation here. Yes when I was freelancing I had EC2 instances running RM Server for clients that I had on a Lambda CRON sched. It would spin up once/day, do its thing, and spin down. MUCH cheaper than leaving it on 24/7.
Has anyone done an apples-for-apples benchmarking analysis using a variety of EC2 instance types with a fairly typical ML modeling task (not DL)? I sometimes feel like I'm just grabbing some config at random sometimes..
Scott
1