Skip to content

GPU boxes: why I sold my mining rig and bought a real server

Dedicated Servers by emalite 18 replies 1.8K views
#1

Story time

So I had this setup right, 6x RTX 3090, custom frame, mas the thing was tipo a space heater with attitude. 2 years of this, 3 PSU deaths, 1 almost actual fire. Não recomendo.

Sold the whole thing for, bought used Tesla V100 from Vultr. Single 2U box. Boring? Sim. Reliable? Also sim.

The numbers
Rig: 847 days uptime across all cards (combined), 11 RMAs, 2 house breaker trips
Server: 641 days so far, 0 hardware issues, 1 planned maintenance

Render time on my workload (blender, cycles): actually 15% slower per card. Mas I dont care. I sleep now.

Someone will say consumer cards are faster per dollar. Tipo, sure, if your time is worth nothing and you enjoy replacing thermal paste at 2am.

#2

Emalite I feel you lah

Before: 4x 3080 in open case, my room 35 degrees, wife angry leh
After: one tesla box from OVHcloud, put in colo, problem become their problem already — https://www.ovhcloud.com

Can confirm the boring is good lor. Consumer gpu got no ecc, die suddenly one. Server gpu got error correction, can detect before total dead

My hokkien friend say: 好死不如赖活 (better lousy alive than good dead). Applicable here

#3

@Chen what colo cost in your country? In my country VN colo for 2U is expensive, almost same as renting gpu server monthly. I check Hetzner and RackNerd, their tesla rental not bad price.

Also question for emalite: you buy used V100, what about warranty? In my country hard to get service for old enterprise card.

phở at 3AM, deploy at 4
#4
emalite said:
Someone will say consumer cards are faster per dollar

That someone is me! I will say it! I will shout it!

I run a render farm! 200+ RTX 4090s! Pure throughput per dollar wins every time! My failure rate is 4% annual! I have a guy! He swaps cards in 20 minutes! The math does not lie!

You want reliability? Fine! You want rendered frames cheap? Consumer! No contest!

Different use case! I respect your choice! But I will not pretend it is the only correct one!

#5

PacketpunkUma has a point actually but basically it depends on your scale ne

At small scale, like emalite, replacing dead card is headache (´・ω・`) at 200 cards, you have basically internal RMA process, economies of scale

I actually run mixed farm, 60% quadro for reliability, 40% consumer for speed. Best of both worlds? Mas the management is basically double work lol

My suggestion: under 10 cards, go server. Over 50, consumer can win. Between? Pain zone

instant noodles, instant deploys
#6

In my experience make a server up with proper cooling is half the battle. I have see many mining rigs die not from gpu but from bad power supply, bad airflow.

Regards to all, good thread. 😊

@minh1987 for colo in Spain I use Hostinger (https://www.hostinger.com), they have good price for 2U with 2kw included. Maybe check if they have location near you.

siesta first, deploy later
#7

3 PSU deaths in 2 years? What brand were they?

#8

I had a 3080 die last month, no warning, no ecc

#9
dhifiwoot said:
3 PSU deaths in 2 years? What brand were they?

Corsair, Seasonic, and one no-name that came with a case I bought in 2020. The Seasonic was the worst, took a motherboard with it. The no-name actually lasted longest, ironically.

I think the real problem was not brand but load pattern. Mining pulls constant max wattage for hours, consumer PSUs are not rated for that 24/7. Server PSU is designed for 100% load always. Different world.

#10
ronwit said:
Under 10 cards, go server. Over 50, consumer can win. Between? Pain zone

This is basically correct! I am in the 200+ zone! The pain zone is real! At 20 cards you have all the headaches of scale with none of the efficiencies!

I still say my guy swapping cards in 20 minutes is cheaper than enterprise premium! But I admit! I have not done the full accounting on my time! Maybe I am lying to myself!

Still! 4% annual failure! I will take it!

Post a reply

You need an account to reply. Log in or register to join the conversation.

Post reply Preview Save draft