EU GPU Deep Learning Server 8xV100 SMX2

Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.

Rand__

Well-Known Member
Mar 6, 2014
6,626
1,767
113
Maybe @boliniak1 has some other options ? ;)

I wouldn't call that a great deal though at 40k GBP but then I have no clue re cost of ML hardware, maybe its a steal?
 

kedzior

Active Member
Mar 21, 2018
121
27
28
49
Poland
Yes, I send PM to him, he is also from PL so maybe we will arrange something.
The cost for this kind of equipment is around 80k-90k USD.
 
  • Like
Reactions: Samir

boliniak1

Member
Sep 28, 2018
57
18
8
Yes, I send PM to him, he is also from PL so maybe we will arrange something.
The cost for this kind of equipment is around 80k-90k USD.
Hello sorry for delay but I was very busy, so to clarify this is build we done for a client and he can't pay for that instead of PO, but this is different story.
Server are brand new and builded from new hardware, main platform deliver from Taiwan.
We only occasionally build 4*V100 systems but if you need we can do it but unfortunately it will be in 4U Rack Case not in 2U case (I don't know your requirements).
 
  • Like
Reactions: Samir

Samir

Post Liker and Deal Hunter Extraordinaire!
Jul 21, 2017
3,257
1,445
113
49
HSV and SFO
Wow, I've never seen torque specs on heatsinks before--very cool! A lot of precision assembly if they're spelling that out. :eek:
 

boliniak1

Member
Sep 28, 2018
57
18
8
Wow, I've never seen torque specs on heatsinks before--very cool! A lot of precision assembly if they're spelling that out. :eek:
Yes to assemble this kind of machine new need to have a lot of o knowledge and few very expensive tools to do it.
 
  • Like
Reactions: Samir

LukeP

Member
Feb 12, 2017
183
21
18
44
or $8K for equivalent power, if you add 8 2080 Ti gpus. why waste tens of thousands of dollars lol (unless someone else is paying of course)

these machines do not make your AI better. its just brute forcing it to fake it
 
  • Like
Reactions: ZXCD

boliniak1

Member
Sep 28, 2018
57
18
8
or $8K for equivalent power, if you add 8 2080 Ti gpus. why waste tens of thousands of dollars lol (unless someone else is paying of course)

these machines do not make your AI better. its just brute forcing it to fake it
OK 8*2080 TI is not the same power (rather 12-14) for the first and second what is more important in Enterprise-Grade Infrastructure you don't have GPU-Direct in RTX.
The last thing in many models 11GB per GPU is not enough.

I also just wondering how you can attach 8 GPU in normal/standard/cheap mainboard - there is difficult to do it even in EPYC architecture.
 
Last edited:
  • Like
Reactions: Samir

LukeP

Member
Feb 12, 2017
183
21
18
44
There are ways around these problems but most people using these aren’t inventing anything new they are just using cookie cutter tensor flow because their department got an ml budget to blow.
 
  • Like
Reactions: Samir

Cixelyn

Researcher
Nov 7, 2018
50
30
18
San Francisco
There are ways around these problems but most people using these aren’t inventing anything new they are just using cookie cutter tensor flow because their department got an ml budget to blow.
If you spend all your engineering time working around these problems, then when do you have time to do actual research?

Also @boliniak1 is right — 11GB doesn’t really cut it these days. Some of the new SOTA models won’t run on anything leas than 16GB (making the Titan RTX your minimum viable consumer GPU, and running those in a dense configuration is nontrivial to say the least)
 
  • Like
Reactions: Samir

boliniak1

Member
Sep 28, 2018
57
18
8
If you spend all your engineering time working around these problems, then when do you have time to do actual research?

Also @boliniak1 is right — 11GB doesn’t really cut it these days. Some of the new SOTA models won’t run on anything leas than 16GB (making the Titan RTX your minimum viable consumer GPU, and running those in a dense configuration is nontrivial to say the least)
Yes, it is absolutely true and we can also point out NVLink speed between GPU cards on standard PCIe GPU vs SMX2.
 
  • Like
Reactions: Samir

LukeP

Member
Feb 12, 2017
183
21
18
44
Yes, it is absolutely true and we can also point out NVLink speed between GPU cards on standard PCIe GPU vs SMX2.
gpu direct and Nvlink are Nvidias way to cover up deep learnings dark secret that it doesn’t scale. it’s not embarrassingly parallel. Researchers are too busy Buying these as a stop gap and not focusing on the hard questions.

this is why I say a 2080ti farm is the go for actual work these days. Because we have moved on from the fad that was deep learning.
 

boliniak1

Member
Sep 28, 2018
57
18
8
gpu direct and Nvlink are Nvidias way to cover up deep learnings dark secret that it doesn’t scale. it’s not embarrassingly parallel. Researchers are too busy Buying these as a stop gap and not focusing on the hard questions.

this is why I say a 2080ti farm is the go for actual work these days. Because we have moved on from the fad that was deep learning.
I'm pretty sure that you did not work with very complicated models as we wrote before 11GB is not enough for the first, the second 2080 TI does not cover all researcher's aspects.
Did you work on this kind of machine and use all hardware potential (8*V100)?
So I just wondering why many Enterprise buys DGX1 or DGX2 maybe they should invest in a big farm of RTX 2080TI servers.
 
  • Like
Reactions: Samir

LukeP

Member
Feb 12, 2017
183
21
18
44
the only people who want you to use big models are the people selling the hardware. its a scam sorry. who did the largest state of the art 33GB model recently? nvidia. hmm what a coincidence. dont be a sheep homey (unless you make money being a sheep and i cant blame u.. :p)
 

boliniak1

Member
Sep 28, 2018
57
18
8
the only people who want you to use big models are the people selling the hardware. its a scam sorry. who did the largest state of the art 33GB model recently? nvidia. hmm what a coincidence. dont be a sheep homey (unless you make money being a sheep and i cant blame u.. :p)
Sorry LukeP but now I'm sure that you did not work with complicated research in DL.
 
  • Like
Reactions: Samir and LukeP