Awesome GPU accelerated instances but UI/docs could use some work
January 16, 2019
Awesome GPU accelerated instances but UI/docs could use some work
Score 7 out of 10
Vetted Review
Verified User
Overall Satisfaction with Google Compute Engine
We use Google Compute Engine for creating cloud virtual machines that are GPU-accelerated. Because we have centralized most of our backend services to DigitalOcean, we use Google Compute Engine for most of the services that DigitalOcean doesn't offer. Additionally, we use the custom image functionality that GCE offers to create VMs with custom images.
- Great scalability. The cloud VMs all have elastic specs functionality, but re-scaling some VMs may create a significant amount of downtime for your backend.
- GPU offerings. Google Cloud offers NVIDIA Tesla K80s, P4s, and P100s, which some of the cloud computing competitors don't offer.
- Downtime, Google's SLA is very good. I've never had a poor experience with downtime or maintenance on their services.
- Internet speed can be quite variable. The bandwidth for different instances ranges a lot. Some instances have had internet bandwidth that is in the range of 5-10x the speed of other instances.
- Customizability. Customizing the number of cores, RAM beyond what Google offers in their standard compute plans can get quite expensive.
- Firewalls/networking. Figuring out how to use these took way longer than necessary. Getting the right ports opened and forwarded took lots of reading, something that other services included in the creation/initialization process of virtual machines.
- Positive. Suggestions on re-configuring specs. If you aren't using all the RAM or CPU threads, suggestions are sent to re-configure your virtual machine to minimize your bill at the end of the month.
- Negative. Expensive custom machines. If you're looking for a more exact figure of RAM or CPUs, the virtual machines get quite expensive fairly quickly.
- Negative. Internet bandwidth variability. The internet bandwidth you get, from my experience, is all dependent on your luck. :(
AWS's UI could use a lot of work, and their API documentation was much worse compared to Google's, which was already tough to read and figure out. Google's free trial of their services through a platform credit (which AWS doesn't offer), also helped us test their compute platform, which helped us figure out how to integrate the GPU-accelerated instances into our workflow for no extra cost.