Making your own server

Awesome. What sort of performance are you getting compared to the AWS basic GPU?

What’s the best way to benchmark the performance?

In the first lesson it took 318 seconds to fine-tune with batch_size=32.
It ran out of memory with batch_size=64

Does anyone have a recommendation on what kind of motherboard I should get if I want to be able to use multiple GPUs (2 for now)?

And is there a specific version of the 1070 I should try to get?


That sounds reasonable. You could compare runtimes / memory issues with our AWS GPUs since we know those are working for most use cases in this course.

most builds I’ve seen use Gigabyte GA-X99-Ultra which would allow you to use up to 4 gpus…
A couple of interesting builds:


MSI and EVGA are the most popular due to performance and warranty. I am a big fan of MSI but EVGA has an excellent warranty.

1 Like

I get 229s with my 1070 on lesson one’s first fit. On AWS P2 I believe it is around 650s.


Setting up your own machine is really easy. Even a 1060 (sub $200) will blow the doors of AWS P2 instance and will be no monthly costs. The 1070 is the sweet spot but will go for around $400. I highly recommend using Linux as Windows is always fighting against the grain when installing dependencies and linux just works faster and smoother. Especially Jupyter notebooks, the difference is huge.


Do I need a powerful CPU to set up a server? Or CPU is not doing any work when training with CUDA?

Consensus is that the CPU doesn’t need to be that powerful, as long as you can support enough processes per GPU (and depending on what kind of preprocessing you need to do).

I’ve seen the Intel 6700 in a bunch of builds so that seems reasonable (and you could probably get by with less if necessary).

1 Like

Here are the components I chose.


Graphics Card





Power Supply



It isn’t CPU bound when training on GPU, but it will peg one core 100%. So single threaded performance does factor in a little. More importantly is the bus, using PCI Express Gen 3 is 10-15% improvement over the same card, and using a faster CPU can be anther 10-15%.

For example, I had a 4.3GHz overclocked Ivy Bridge (Intel 3770K) with an nVidia 1070 and would get around 324s times training lesson 1 first fit. When I upgraded to a Kaby Lake 7700k using the same GPU but faster CPU and Gen 3 PCI Express, my times dropped nearly 30% to 229s.

When I train on GPU I can see one thread at 100% usage. So CPU does factor in a significant amount with GPU training but you will already be running quite fast, it’s just the difference between fast and faster and no where near as important as the GPU in general.

1 Like

Good choices, I would recommend going Kaby Lake though and if you can just spend the extra $100 on the 7700K. You can also get a Z270 which is the newer Intel chipset for near the same money.

I’d also recommend swapping out the Power Supply for an EVGA Gold G2, it’s about $10 less but it is a much better Power supply and has a $20 rebate (just got one myself).

1 Like

Good call. I’ll take a look.

What are your training times per epoch for cats and dogs on this configuration?

For lesson one, it is 229s to do one epoch on the first fit. On AWS P2 it is around 650s.
I am using the following:

Kaby Lake 7700K
MSI Z270 M5
MSI nVidia 1070 8GB Gamer X
32GB Ram

1 Like

@jeremy I want to suggest that you do this people a mail so we can have gpu access, since its for learning purpose.

dradientgescent –

I’m running two different Linux servers but am not achieving the same speeds as Jeremy does in his examples.

One is a newer machine that came equipped with a 1060, the other is an ancient (was running Vista) machine that I gave a used 1070 (like $20 from eBay). Both are running Ubuntu 16.04, but both run the notebooks slower than the video.

Are you getting speeds comparable to the speeds Jeremy is getting in his videos?


I didn’t pay attention to the times Jeremy got. I just know the times I get.

But the 1070 will be slower on a gen 2 pci express port or a slower CPU. Will still be a lot faster than AWS p2 instance but won’t run at full speed unless on a x16 PCI Express 3 port and Kaby lake CPU.

1 Like


Looks like a decent rig, few suggestions: Get a platinum+ power supply, I like seasonic for performance vs price. It’ll pay you back for itself eventually, no reason not to get. Things to note about power supplies is they deliver their most efficient power @ a certain range of wattage, check out some spec sheets. I’m not familiar with that fan, but I’d get a noctua if I was on air almost period. They’re silent, very effective, 6yr warranty, and generally ugly and brown. I’d get the 7700k, you want the extra PCI lanes, and various other goodies. I’d get almost another other branded card besides zotac. Asus is the best with the 1080 imo. I might also hold off looking for what the prices are going to be for the 1080ti if I was buying now.

1 Like