Same story here. 3070, gaming load not even mining. Black screen, dead. No logs, no SMART for GPU, just gone.
Now I rent A40 from Vultr Warsaw for my ML experiments. Price hurts compared to owning, but I can sleep through the night. I use https://www.vultr.com/pricing/ when I need to remind myself why.
Member
Thanks but Hostinger no colo in VN region, closest is Singapore. Also I check their site, they mostly sell shared and VPS, colo not main business. I think you mean their cloud?
For VN local, I found Vultr has Singapore DC which is okay latency. Hetzner no Asia DC except Singapore also. RackNerd only US/EU.
Actually I think I will just rent GPU instance monthly, forget colo.
Has anyone here actually priced colo in Zurich? I checked OVHcloud Limburg and Hetzner Falkenstein. With the CHF/EUR spread and transport cost, renting wins unless you have unusual power density needs.
I run 2x A100 in OVHcloud Beauharnois via their US pricing. Works out fine for my BGP experiments.
Member
No no, I mean colo. But you are right, I check now and Hostinger colo is only in some locations, not general offer. Maybe I got special deal or old plan. Sorry for confusion.
In Spain I know people use Interxion or Digital Realty for real colo. Hostinger was exception, not rule.
Pain zone is right lor. I was at 8 cards before, exactly there. Not enough to hire staff, too many to hand-fix. That is why I jump straight to colo with OVHcloud Singapore, let them handle — https://www.ovhcloud.com
OVHcloud Singapore DC is real, I visit once. Hot aisle cold aisle proper, not my bedroom with box fan lol.
This is not related to thread but I feel you spiritually
Back to topic: I am curious, anyone try the new cloud GPU from Vultr? They have A100 now I think, not just old Tesla. Price is... not cheap, but per hour good for burst render.
Following this thread. I am in the pain zone with 12x 3090, trying to decide same thing.
This is why I tell people: buy the PSU warranty that covers connected equipment, or don't bother. Most don't.
Also re: V100 being slower per card, have you checked if your Blender version has proper CUDA path for Volta? Some builds are weirdly optimized for Ampere and newer. You might be leaving 5-10% on the table.
12 is basically the worst number. Too big to baby, too small to automate. My condolences.
Interesting. I use Blender 3.6 LTS, I think Volta is supported but maybe not optimized path. I will check the build flags, thanks. Even 5% back would be nice, though honestly I am not going back to consumer cards either way. The sleep factor is the real metric.
