Light duty server with gpu for ml

[jeffbaumgardt]

Budget (including currency): $1,000 - $1,400 USD

Country: United States

Games, programs or workloads that it will be used for: Server application, docker with 2 - 3 containers. 1 gpu ML. Other containers will be very light duty.

Other details (existing parts lists, whether any peripherals are needed, what you're upgrading from, when you're going to buy, what resolution and refresh rate you want to play at, etc): I'm building from scratch, only need box and internals, peripherals are already in house.

 

I know with a budget is $1,400 is very tight considering a gpu. I can bump the budget by another $200 for the right setup. I prefer to start with 8 - 16gb ram but as a single dimm as I will want to be able to upgrade more later.

 

Hard drives are not as important, maybe 1 - 2 tb is more than enough. Not doing anything that needs high bandwidth for HD read/write.

 

The primary use case for this is to run a ML processing container that will be running against a 60fps onvif camera. I have both cpu and gpu use cases but the models for the CPU are much less so I won't be making grilled cheese on the CPU.

 

Cooling doesn't need to be over the top, maybe an aio closed loop rad for the CPU but that's the extent, need plenty space to add more cooling as needed just in case. Accustics are a minor consideration as this box will be in my office with me so it shouldn't sound like a jet plane. But it's far from important.

 

[Radium_Angel]

13 minutes ago, jeffbaumgardt said:

important

Have you considered a refurbished workstation? Something like a Dell Precision T-line, they come with dual CPU Xeons, tons of RAM, and a good (non-gaming) video card like a quadro, which might work well for your ML needs

[jeffbaumgardt]

That's a really good idea. Would I be able to swap out the Quadro for a 1080 perhaps? Or would there not be enough pci lanes? I don't know too much about that line so I'd have to research what the Mobo has.

[jeffbaumgardt]

I haven't benchmarked a Quadro with my ml so I don't know if it's any good or not.

[jeffbaumgardt]

From a cursory read Quadro isn't great for "scientific" needs, so probably not great for ml. But GeForce I have benchmarked and ran really good. So I just need to find out now if I can swap that gpu out.

[Zando_]

8 minutes ago, jeffbaumgardt said:

That's a really good idea. Would I be able to swap out the Quadro for a 1080 perhaps? Or would there not be enough pci lanes? I don't know too much about that line so I'd have to research what the Mobo has.

PCIe lanes would be fine, if you pull a GPU and replace it with another one, the amount in use stays the same. If the board has PCIe 2.0, at x16 you'd need a 2080 Ti before you'd start getting bottlenecked. PCIe 3.0 x16 will still handle anything on the market, and only the 2080 Ti loses 3% or so when dropped to x8, everything below that would be fine. 

1 minute ago, jeffbaumgardt said:

From a cursory read Quadro isn't great for "scientific" needs, so probably not great for ml. But GeForce I have benchmarked and ran really good. So I just need to find out now if I can swap that gpu out.

Should be able to so long as it's a full sized tower, just make sure the PSU has the connectors you need for the 1080 you have in mind (some want one 8 pin, others want 2). 

Also if you quote people like I just quoted you (arrow underneath their post), or tag them like so: @Zando Bob, it'll give them a notification that you replied. 

[jeffbaumgardt]

@Zando Bob thank you so much. I've built gaming systems before in the past but all parts were brand new so I didn't worry to much about compatibility.

 

This looks like a really good option. Thank you again.