Star Nursery: Early but fascinating days for the new era of GPU cloud compute [BLOG]
Published: September 1st 2016
As we begin to grind away in the cloud with testing, processing, and even more testing, it is still somewhat unclear what the best approach will be for selecting big cloud provider compute. That said, I’m happy to share some of the knowledge we have gained in the past year as we have worked to stand these solutions up with our partners.
The first location we began testing was the obvious one – AWS, using the EC2, g2.8xlarge, their biggest and best offering for GPU resources (Cinnafilm Image Processing saturates the GPU at 99% on average, meaning we can never get enough cycles – we feed on speed). Unfortunately, Amazon’s best foot forward today is the K520 card; a decent gaming card but running at a sluggish N-2 generation speed. These cards will technically operate our software, but slower than we prefer since we recommend a Tesla-series card at a bare minimum, to even provide tech support. Given how far NVIDIA has come in the past two years since the K520, it is surprising that Amazon has not made (publicly known) any plans to migrate to more modern GPUs that can really keep up with the types of needs we will be exploiting in the cloud. Overall opinion: AWS is stable, overpriced, but too slow for our recommendation. Definitely not powerful enough to process anything in UHD.
The next place we are looking is Azure. On the heels of AWS’ announcement of Elemental last year, Microsoft boldly announced it too was heading into the GPU space. It has been a quiet year since, and we were beginning to lose hope until this very recent announcement just a few weeks back. Needless to say, we are VERY excited about this. However, “preview availability” means select entry and some time before they are readily available to the masses. That said, we are really hopeful that this solution will hit the market soon, as we highly recommend these GPUs to everyone we know. Overall opinion: looks great on paper, can’t wait to try it.
There are rumors Google may be working on a solution, the question is – are all of their GPU cycles going to focus on deep learning and AI, or will they stand up a solution that is tuned for video as well. Overall opinion: We will cross our fingers for some news soon.
SoftLayer is our bare-metal cloud of choice, though their pricing model is rough. They bill at a monthly rate and the prices go up quickly with pretty standard options. Overall opinion: Very solid choice for longer-term, steady offsite usage. Elasticity limitations, a bit pricey (with some up-charge options not very clear). Some stand up latency. Highly supportive of small businesses.
Cisco is another company that has made some hints to GPU cloud processing, but we have no other information besides hints at this time. Overall opinion: nothing of note.
It is obvious: the company with the best view to this vast, new star nursery is NVIDIA. They are most definitely in the driver’s seat, and with the newest M60 Tesla cards coming online (running at another ~40% speed increase over the very reliable and stable K40), it is safe to say that their future is looking very bright indeed. Companies that are adopting these incredible NVIDIA GPUs in these business solutions are able to take advantage of massive compute power, a necessary ingredient in managing exponentially increasing video processing workloads.
Basically, these are pretty early days when you talk to the cloud providers. Except… they are not, really. Cloud and virtualization was a hot-button topic at NAB 2014, and then everyone started focusing on UHD, VR, HDR, etc… There really is no reason the cloud providers can’t go all in on GPU cloud compute – I wonder if they have some of the same distractions we do in M&E; that is, they are only focused on AI/VR/ etc. and not seeing the massive volume of video number crunching as an opportunity. At least Amazon has the right idea, and Azure looks like they get it now – we can only hope they get their rocket off the ground before something else catches their attention.
Parting note – Dark Energy will eventually migrate to the cloud as well – look for related announcements toward the end of the year starting at SMPTE in October. Come see us at booth 120!
— Lance Maurer, CEO