Not that long ago, most cloud providers offered solutions for capacity planning, scaled CPU computation, virtualization and security. GPU specific task based solutions however were sparse.
For small business or the consumer, they were nearly non-existent.
Recently, GPU clouds have emerged, offering what could be the most sensible use of PaaS and SaaS cloud usage models. Established giants and emerging startups are selling fully packaged cloud solutions, with the networking, hardware and OS provided for you. Some offer specialized applications fine tuned to get the most out of a GPU.
Consider this scenario. You work for a mid size engineering company, albeit one with a tight budget. As a part of your research, you simulate things like fluid dynamics, stress physics and other related characteristics of your products.
You need to be able to calculate the effects of the slightest subtle changes and visualize them.
Without the aid of clustered workloads in the cloud, its up to your IT department to provide the GPU power needed for such high demand computing. Depending on the size of your team alone, it can cost upwards of $2k to $5k per seat. As demand for faster responses to calculating change rises, so will the demand to upgrade your infrastructure.
Filming and game studios encounter a related challenge. The demands placed on them for realism in 3D objects translates to big costs of both money and render time. To stay competitive, they continuously improve on techniques and added effects.
Such cost and performance conundrums are what make GPU clouds ideal. Instead of spending precious capital for the latest hardware on site (which will inevitably become obsolete well before you see decent returns on the investment), you can rent the graphic processing power when you need it.
In addition, when the GPU resources are idle, you aren’t wasting money.
GPU clouds are even more sensible when you consider the underlying technology (they were kind of made for it). Unlike standard CPUs, GPUs excel in parallel processing of many tasks (thousands) at once. Using this feature, providers can offer shared options that take maximum advantage of process threads and runtime. In some use cases, single GPU instances can supplement the workloads of racks of other VMs. Why lock up all those stream processors and compute units to one user workstation when you can maximize use for the team?
Nvidia and AMD each offer their own solutions via Amazon EC2, Appstream, Google Cloud Platform, Microsoft Azure and other providers.
Appstream in particular is a very promising solution, offering 3 Graphics centric instances to choose from, each supporting the standard DirectX, OpenGL, OpenCL and CUDA libraries. These instances are most suited for generalized parallel processing, for fields like data analysis and machine learning.
Paperspace is another interesting company, offering their Gradient suite of tools with simplified GPU job management using simple python commands. These solutions are mainly targeted for research start-ups and enterprise clients. For entertainment, the average consumer wouldn’t see a cost benefit. Thankfully, there are options in that category too.
Nvidia’s GEForce Now is a service offering access to GEForce GPUs in the cloud from any thin client. They support hundreds of games on the platform, with more to be added.
Shadow is another consumer cloud product tailored to gamers, but offers machines for more general use as well. These are powerful machines running GTX1080 equivalent GPUs. Fitting well into the PaaS category, Shadow PCs are server grade machines, running windows 10. They tout an “always upgraded and always updated” pitch, so users just have to worry about their own apps and data on the machines. They offer fiber network connection options and 4K streams.
The subscription price at the time of this writing is $34.95 per month ($419.40/year). This seems high at first glance, but when you consider building or purchasing an on premise PC with the same specs (near to what they offer), and you’re looking at a value equal to over 3 years of Shadow subcription service. (I built a rig on PC Part Picker, picking the cheapest parts I could put together that still matched Shadow’s advertised specs, and the total came to $1354.16. I’m talking the lowest quality parts here.)
The price for the hardware may be on point, but also consider the usage charges for going above thresholds. If you work for a large firm, consider their backup and recovery options.
Security, Security, Security (and Security too):
A GPU cloud solution may be the perfect answer for machine learning and 4K AAA gaming alike. Before you start frantically subscribing, its important to consider the major security implications of transitioning to the cloud paradigm.
A good solution might offer traditional hypervisor controlled VMs or application specific containers. In either case, its important to research the provider. Seek to ensure your proprietary work is protected via isolation from other instances. As most providers fit into the public cloud category, dedicated virtual private servers are preferred.
Any connection to your GPU workloads should be encrypted via PKI, with a seamless certificate update/revocation scheme, and void of obsolete or broken algorithms. Direct connections to sensitive GPU output should be provided via encrypted tunnels (L2TP with IPSec or better). If the application is browser/web based, grand “old” HTTPS via TLS is preferred (even with VPN tunneling in the mix). A layered security approach, focused on the network and application levels is paramount. If your provider is unwilling to at least guarantee that similar security standards are in place, its probably time to look into other service options.
If you’re a bold one, you might even inquire about your providers pen testing policy. Are they regularly validating their security?
GPU clouds are up and coming, with an already great lineup of options for you to choose from.
Maybe you’re already using one or have experience with them. I would love to hear your input on this and related cloud tech in the comments.