Just seeing if anyone has found good results with running CR in the cloud.
After a few hours of banging my head against the wall I got a MS Server 2016 instance on a g2.8large, which has 4x GPUs (know RC only binds to 3 GPUs without extra licenses) with some serious cuda cores and 32 CPUs. Took a while to find recent drivers but got the proper drivers loaded up as registered K520s . I then threw about 5GB of images from AWS S3 onto the SSD drives and let RC attack a normal reconstruction.
Performance was about par with my rendering machine at home (specs: 32GB RAM, Nvidia 980 ti, intel 4600k, NVMe storage), maybe a little faster given that I know ‘normal reconstruction’ now is about 20% ‘slower/better’ after an update? I was tremendously underwhelmed however - definitely not worth the money to render in the cloud if I got everything right. I may build another machine to double check on the p2.8x instance (faster GPUs, though a bit more expensive) if I get any contrary responses in this thread…
Has anyone seen insane speeds from EC2? For those that have upgraded their machines to some quadros and xeons, what is the improvement over a ‘consumer’ rig? I was running 2x Nvidia 980tis at one time but for other applications saw only about 20% improvement as the bottleneck seems to be the CPU. The CPUs were pegged during my test render on the EC2 instance but again, it doesn’t seem any faster than my home rig, just more expensive.
I have a few reasons to want to render on EC2, namely because I need to put a lot of images from different places on S3 to throw them at CR and then archive the images onto Amazon Glacier. Pulling the images from S3 back to wherever my rendering machines are equates to some very fluctuated bandwidth costs I would like to avoid.
While I may have to invest in that regardless, I’m also thinking of going another route and with my next build (if I get some business), and go for one of those 10 core intels on a z170 board with a 1080 ti or two once they are released. There would have to be some significant improvement over the same 4 core CPU / single GPU build though and I’m just not seeing it from my experiments.
Hi Castlenock
the main problem here is a VIRTUALIZED hardware and, if I remember correctly, the AC2 instances are 15-25% slower compared to the same hardware on your desk running without virtualization.
The Tesla GPUs are not the fastest ones, gaming GPU is faster by 20+% as it uses much higher clocks…
Try to look for a lower CPU core version with the same GPU count as too many cores bring no performance anyway… There are the Workstation CPUs that are aimed at performance, lower core count but much higher GHz…
While I may have to invest in that regardless, I’m also thinking of going another route and with my next build (if I get some business), and go for one of those 10 core intels on a z170 board with a 1080 ti or two once they are released. There would have to be some significant improvement over the same 4 core CPU / single GPU build though and I’m just not seeing it from my experiments.
A local fast PC will be always faster than a cloud solution… at least in the foreseeable future…
Thanks for the quick reply! What you say makes sense - I still may need to do some crunching in the cloud as again, I have some weird workflow things I’m considering which means a could have a lot of cloud storage that would cost about as much to pull down to my local machines as it does to pay for a spot instance.
Also, it’s handy to render something remotely as my main rigs don’t always have a great Internet connection or is available for reconstruction. So that it works at all is a good option, and the steam client makes switching between my home rig and an EC2 instance possible.
Could you clarify this for me?
Wishgranter wrote:
Tr look for lower CPU cores version with same GPU count as too many cores bring no performance anyway… There are the Workstation CPUs that are aimed on performance, lower core count but much higher GHz…
I’m doing whatever I can to keep things as compact as possible for my desktops: having a consumer motherboard (Z170) lets me get a lian li case with a handle which makes my machines somewhat portable and still have a great deal of punch and flexibility. That and it’s cheaper when I add up a xeon rig that incorporates NVMe and other stuff - so I’m just figuring out if my next processor is going to be best for CR:
For the US, I’m looking at:
Intel Core i7-6700K 8M Skylake Quad-Core 4.0 GHz at around 340 USD
or
Intel Core i7-6800K 15M Broadwell-E 6-Core 3.4 GHz at around 440 USD
I’m confident the last isn’t worth it of course… Still what would you guess the improvement would be between 1, 2, and 3? If any? If I read Nvidia’s stance on things to come, they’re backing off SLI for a variety of reasons, so I’m planning on one big/beefy GPU rather than multiple GPUs.
Hi, I have recently tried using a G3.16XLARGE instance in Amazon AWS EC2 to process a job (4x Tesla M60 GPU’s, 64vCPU and 488GB RAM) and it was no quicker that a 3x GPU card workstation. Interestingly in the cloud watch monitoring tool I could see it maxing the CPU usage. It’s a shame there’s no simply way of monitoring the GPU performance. I find it hard to believe that this isn’t any quicker.
Has anyone else seen similar results? Perhaps someone has looked at the Elastic GPU product?