400T - “400T total, 6 node cluster with 60T of NVME is around $9k per month”
3 Likes
Actually, the advantage is the opposite, to go from 100 to 0 from a flick of a switch, which means a lot if you don’t have a couple of million to spare.
Regarding the best ‘pricing calculator’, yes you are right, but we can’t compare apples with oranges, meaning, our latest job had a ton of AI work which as enourmous implications on filesystem and throughput, if we compare that against a 1 flame on-prem job it will look ridiculous.
But to put it in context, perhaps it helps to check a job we did in March with 2 flames using g5.12x instances, 8TB FSx and 4Tb x2 EBS volumes for the flames with 1024Mbps throughput on FSx (if I recall well)
Total cost of everything, including license servers, production servers, production machines as well, OKTA servers…
Also, 1 CG artist using the renderfarm (not much, only 15 machines) for 2 weeks.
USD 4,063.68
1 Like
Hi Jordi,
Thank you providing a cost breakdown. And I totally see what you mean in regards to sporadic allocations. That price isn’t bad if you just need something for 2-3 weeks.
What I find interesting is your choice of instance type. A g5.12xlarge has 4 GPUs with with 24GB of VRAM each. Flame only uses a single GPU, so you basically have 3 GPUs sitting there idle , and still only using 24GB of ram. Now maybe the deep fake stuff you were running can allocate of GPUs though? For Flame, something like a g5.16xlarge seem more appropriate.
But, the end goal of any business is not to have a job here and there for 2 weeks, rather to be earning revenue on a consistent basis, and that is were cloud just doesn’t make sense. You spend more money on less capable hardware with increased complexity.
Thank you for the dialog.
Absolutely. There is a variation though - where you are working consistently, but not all your work is on Flame. Other days you maybe on Nuke, Avid, Nuendo, C4D, etc. There are businesses that scale in a single tool and a narrow job scope. There are businesses that have a different mix of work. No right/wrong there. But don’t conclude that short-term needs equate lack of work.
In fact the cloud could be useful in that scenario as you can tailor your systems to these specific needs, or keep certain needs on-prem while having other peak loads in the cloud when the need exists.
Thanks Alan, I will check the 16x, we have been quite successful with the 12x, perhaps is due to its network bandwidth but I see what you mean with the GPU… strange Flame does not support multi-GPU setups.
Indeed, in a way you buy flexibility to the Nth degree and in fact, although you want projects moving non-stop in an ideal world, well, that does not happen all the time so actually AWS makes even more sense when you don’t have work at all becuase you stop the costs of your hardward and you don’t have any debt.
I think the issue will be when we are so busy that financially it questions the whole model… let’s see.