I challenge Linus to build a cheap Data science PC

[Pensive]

I came into this site when I was looking for a data science computer for personal use , and found that their most expensive build is around 50k $ , then i wondered how it would hold itself against the maxed out mac pro , then i wondered if Linus can build a better performing PC for half the price

 

The one rule I have for Linus in this challenge is that he uses real world machine learning challenges for benchmark , not a benchmark app

The site that sells data science PCs : https://www.pugetsystems.com/recommended/NVIDIA-TITAN-RTX-Data-Science-PC-231

Are they the Apple of Windows world ?!!

[RasmusDC]

if you begin putting titans into a pc, then it becomes expensive quite fast..

 

i would not define puget as apple of pc´s since their focus is performance, Apples is more, Design, simplicity and also a strict software setup. 

 

it is two completely different approaches... and of course if you do "configure" an apple Mac pro pc at a larger vendor like Dell, or other, then you end in the same price bracket, since "professional" workstations are expensive, the margins are higher since volume is lower, and most of the hardware are custom.

[Pensive]

3 minutes ago, RasmusDC said:

if you begin putting titans into a pc, then it becomes expensive quite fast..

 

i would not define puget as apple of pc´s since their focus is performance, Apples is more, Design, simplicity and also a strict software setup. 

 

it is two completely different approaches... and of course if you do "configure" an apple Mac pro pc at a larger vendor like Dell, or other, then you end in the same price bracket, since "professional" workstations are expensive, the margins are higher since volume is lower, and most of the hardware are custom.

Such nice points .

 

I don't know how to choose the best performing GPU according to value properly for AI and Machine learning , and there aren't a lot of videos on YouTube comparing different building for these tasks ,, that's why l thought immediately about Linus ,, he loves to prove he can do better build for lower price

 

 

Software wise , I personally became a fan of Linux since it is free , doesn't slow overtime and forces you into being technical .

 

[GoldenLag]

1 minute ago, Pensive said:

don't know how to choose the best performing GPU according to value properly for AI and Machine learning

Best is just to get as much performance per GPU as possible within your budget. 

 

• Get a mining shelf.
• Get a Epyc CPU for lanes (1k$ including board and CPU)
• Get extencion cables and adapters
• Then have fun adapting untill you have the desired number of lanes per card
• Hope everything works. 

 

The truth is that you can get stupidly powerful hardware if you want to do machine learning. 

 

If you want to do more things, then there are more balanced solution. 

 

But the Epyc setup i described above would be about 2-3k spent in supporting hardware, and rest spent on GPUs. Offcourse people who deliver actual Pro hardware has a lot of support surrounding the services. 

[Enderman]

Literally any PC can be used for data science.

It just depends on how well it performs.

 

This is just another "build a PC for cheaper than the mac pro" topic.

[Pensive]

1 hour ago, Enderman said:

Literally any PC can be used for data science.

It just depends on how well it performs.

 

This is just another "build a PC for cheaper than the mac pro" topic.

But would you rather wait 3 days to run a simulation or wait 3 hours ???

[Enderman]

3 minutes ago, Pensive said:

But would you rather wait 3 days to run a simulation or wait 3 hours ???

The point is you can always spend more money and get something faster.

So it comes down to how much money you make doing your job and what you can afford to spend, not "does this outperform a mac pro"

[igormp]

1 hour ago, Pensive said:

I came into this site when I was looking for a data science computer for personal use , and found that their most expensive build is around 50k $ , then i wondered how it would hold itself against the maxed out mac pro , then i wondered if Linus can build a better performing PC for half the price

 

The one rule I have for Linus in this challenge is that he uses real world machine learning challenges for benchmark , not a benchmark app

The site that sells data science PCs : https://www.pugetsystems.com/recommended/NVIDIA-TITAN-RTX-Data-Science-PC-231

Are they the Apple of Windows world ?!!

 

That PC is expensive because Puget went through the hassle to build it for you and preinstall all of the most used software for ML and DS. A guy made a nice video on it. If you were to build the exact same PC, it would be somewhat cheaper already, since it uses regular off-the-shelf components.

 

The Mac Pro is not a machine for ML since it has AMD cards, hence no CUDA at all to work on anything. Even my crappy machine would be better than it.

 

1 hour ago, Pensive said:

Such nice points .

 

I don't know how to choose the best performing GPU according to value properly for AI and Machine learning , and there aren't a lot of videos on YouTube comparing different building for these tasks ,, that's why l thought immediately about Linus ,, he loves to prove he can do better build for lower price

 

 

Software wise , I personally became a fan of Linux since it is free , doesn't slow overtime and forces you into being technical .

 

You don't even need a GPU if you're just getting into ML. Most stuff can be run off your CPU without much hassle. If you opt to go with a GPU, AMD is worthless in this space, and you GPU becomes useless as well since everything you're doing will be offloaded to the GPU. I used to use a 1050ti as a stopgap after giving away my rx480 and buying my current 2060S, and it served me just fine.

 

I can't really recommend a specific GPU for you since it depends on your workload. RAM might be more valuable than raw power, as an example. Since I mostly work with CNNs and MLPs, I went for the cheapest card with 8gb of vram and tensor cores, and it already gave me double the performance of Colab's P100. (see pic bellow), even though I have an old FX6300.

 

 

Speaking of colab, it's an amazing way for someone to start with, since it already has everything already installed, a nice jupyter environment, and a free GPU! No need to have an expensive PC for that. Heck, I used to train models on it using my chromebook without any problems.

[RasmusDC]

26 minutes ago, Enderman said:

The point is you can always spend more money and get something faster.

So it comes down to how much money you make doing your job and what you can afford to spend, not "does this outperform a mac pro"

that is where i am a bit curios of what actually is the best solution, in the older days we build wolfpacks, a pack of pc´s that shared the calculations or renderings between them, it actually worked extremely great, so you could build quite a powerful wolfpack without the use of extreme amounts of money and then just run the GUI on you´re main pc, we used it for rendering in 3d studio, if we rendered video, which was where it is easiest since you just push a frame for each pc, and then the next when the pc requests it, one pc might be slow but that dosen´t really matter..

 

but when we talk data science, normally in the companies i have worked in, the user interface is the only thing you have on your local computer, the rest is based in server farms, so cost of local pc is low. 

 

even privatly a simple rack mounted calculation unit, might be a better solution that having it all in your desktop pc.

[igormp]

6 minutes ago, RasmusDC said:

that is where i am a bit curios of what actually is the best solution, in the older days we build wolfpacks, a pack of pc´s that shared the calculations or renderings between them, it actually worked extremely great, so you could build quite a powerful wolfpack without the use of extreme amounts of money and then just run the GUI on you´re main pc, we used it for rendering in 3d studio, if we rendered video, which was where it is easiest since you just push a frame for each pc, and then the next when the pc requests it, one pc might be slow but that dosen´t really matter..

 

but when we talk data science, normally in the companies i have worked in, the user interface is the only thing you have on your local computer, the rest is based in server farms, so cost of local pc is low. 

 

even privatly a simple rack mounted calculation unit, might be a better solution that having it all in your desktop pc.

For actual deploys, usually companies rely on cloud services or local servers. In my case, the company I work on uses GCP for everything, so we usually train our models locally whenever possible, or make use of a VM with a couple V100s. After having our trained model, we deploy it into Cloud Run (managed container service from GCP) or into a GKE cluster with TPUs/GPUs, depending on the workload.

 

For a hobbyist just getting into that kind of stuff, there's no need to spend horrendous amounts of money, it's better to rely on freely available services such as Google's Colab or Kaggle, and get by with what you have until you can get money out of it.

[MoVo]

8 hours ago, Pensive said:

But would you rather wait 3 days to run a simulation or wait 3 hours ???

It doesn't really matter. You're not going to run a full simulation or a full training on your local machine anyways, at least not, if your doing this professionally. My company for example usually uses Azure GPU instances for training ML models. They have a lot more horsepower than the individual machines of the employees and can be shared between them, this makes them a lot more cost efficient than buying a huge workstation for each employee. Most people in our department actually don't even have a very powerful notebook, since all the resource intensive processing is offloaded to the cloud.