Today’s article is a direct response to a viewer of this channel who asked for some help in completing their build for their new Deep Learning PC.
The viewer is the lucky owner of an RTX 3090! To get this graphics card, in the current market, it’s a steep investment! When you spend this much on a graphics card alone, you have already blown the budget.
Is the RTX3090 still the best option for a Deep Learning PC?
The RTX 3090 is a “bargain”, at roughly half the price, when compared to professional GPUs like the Nvidia A6000. Great for gaming and also great for professional work.
Before we continue, let me just discuss the new elephants in the room.
Let’s take a quick look at the overall specs:
|GPU Specs||RTX 3060||RTX 3060ti||RTX 3070||RTX 3070ti||RTX 3080||RTX 3080ti||RTX 3090|
|Memory Interface Width||192-bit||256-bit||256-bit||256-bit||320-bit||384-bit||384-bit|
|Memory Bandwidth||360.0 GB/s||448.0 GB/s||448.0 GB/s||608.3 GB/s||760 GB/s||912 GB/s||936 GB/s|
|Recommended Power Supply||550W||600W||650W||750W||750W||750W||750W|
In terms of GPU memory, the RTX 3090 comes out on top, with 24GB of GPU memory. And the RTX 3080ti comes next with only 12GB of memory.
It also has the highest number of CUDA cores, 10496 and in terms of memory, bandwidth is the fastest with 936 GB/S.
When building a PC, it is important to think ahead in how we are going to use our PC.
Let’s be honest, most of us enjoy gaming. So any of these graphics cards will do the job very well.
But if you are really serious about gaming perhaps only the RTX 3080ti or RTX 3090 will do.
If you are doing video editing or even 3D modeling and animation, having a powerful GPU definitely helps. The more memory and CUDA cores, the better.
What about the requirements for Deep Learning?
It really depends on what level of Deep Learning you are at. If you are a Beginner and are on a budget I recommend the RTX 3060 graphics card. It gives you 12GB of GPU memory, which is only less than the RTX 3090. Yes, the memory bandwidth is a lot lower. But chances are, as a Beginner, you don’t have to worry about that. The more GPU memory you have, the better, because that is the biggest limiting factor when trying to run training jobs on large datasets.
But then again, if you have already bought an RTX 3090, then the budget is not an issue!
Then this is the article for you.
Let’s start our shopping trip!
Picking the components
I will be picking the parts for the Deep Learning PC. I will put in our build some RGB lighting. Perhaps the only place I will not bother about RGB is the power supply.
The RTX 3090 graphics card is big, so we need to make sure that we pick a big enough PC case. My recommendation is the Phanteks Eclipse P600S. This case is ideal for cable management, especially for someone who never did it before. It is definitely big enough for a graphics card like the RTX 3090.
With such a powerful graphics card, it is important that the CPU is not going to limit the power of the GPU.
We have a choice of Intel or AMD.
This is a no-brainer nowadays, AMD every time.
In this build, I will go with an AMD Ryzen 5900X as a minimum. I personally own the AMD Ryzen 5900X and its a great CPU, with 12 cores and 24 threads.
But I can totally understand if you decide to go with the AMD Ryzen 5950X, which has 16 cores. But here we are having diminishing returns as there is a very significant uplift in price.
For the motherboard, we need to make sure that it is compatible with the AMD Ryzen 5000 CPU. There are a few chipsets we can go with, for example, the X570, B550, X470, or the B450.
I will go with the X570 as it’s the most recent chipset and also because it has support for PCIe 4.0. Yes, the B550 supports it too. If we were on a smaller budget, it would be possible to compromise here
USB Bios flashback Support
Another important factor in choosing the motherboard is whether it supports USB Bios flashback. USB Bios flashback allows you to upgrade the motherboard bios without requiring a CPU to be installed. This is particularly useful especially when your CPU is not compatible with the BIOS.
Also, if a BIOS upgrade goes wrong, you can easily roll it back. Without USB Bios flashback, life gets a lot more complicated!
Don’t forget the extra USB-C port that you will need
And another small detail that you might miss. Our PC case has a USB C port in the front of the case. We want a motherboard that has a USB-C port header that will connect to the USB-C port in the front of the case.
Most of the motherboards will have a USB-C port in the back of the case, but that’s inconvenient.
So after all these requirements, the motherboard I recommend to you is the same one as I have. The ASUS ROG Strix X570-E Gaming ATX Motherboard. The bonus is that this motherboard also comes with built-in wifi. One less component to buy.
We want at least 32GB of memory, for the Ryzen 5000, DDR4 3600MHz is the sweet spot for memory. We want RGB, so let’s go with the:
This is an important factor. 750W is the minimum power recommended for the RTX 3090. The RTX 3090 itself draws 350W of power. To give a bit of slack, I will go with an 850W power supply. If you plan to have more than one GPU, then you will need a bigger power supply. But my assumption here is that one is enough.
CPU Air Cooling vs CPU Liquid Cooling
This is a hotly debated topic. Is it better to have CPU air cooling or CPU liquid cooling? Both are valid.
But there are some concerns that I have. If we go with CPU air cooling, it is not going to look so good. The CPU fan will also be an obstacle and take up more space in the case. When things get crowded, the air circulation will also be affected.
Also in the future, you will probably want to replace some components in the PC after you do your build. For example, add more memory. Chances are then that you will have to remove the PC fan first.
Another consideration is noise. I personally prefer to have a silent PC, and liquid cooling is the accepted way to achieve that.
So let’s pick an All in One(AIO) CPU liquid cooling.
There are different sizes of liquid cooling: 120mm, 240 mm and 360 mm. The size refers to the radiator itself. The only limiting factor is the size of the case. So you should pick the biggest radiator that you can fit in the case. For an RTX 3090 build, 360mm is the right size.
There are plenty of options.
For my own PC I bought the Corsair iCUE H150i PRO RGB Liquid CPU Cooler (360mm)
This is a good AIO liquid cooler, at a good price, and a good option if you are looking to spend less.
Today I would buy instead the NZXT Kraken Z73. I prefer it because it can display the CPU temperature on an LCD screen. Being able to see the CPU temperature easily that’s quite useful.
As a minimum, you should be getting a solid-state drive. There are two main types of SSDs: The SATA SSD and the M.2 SSD.
The SATA SSD is a drive that is connected to the motherboard via a SATA interface.
On the other hand, the M.2 is connected directly to the motherboard using the PCIe interface.
What is the difference? A lot, speed-wise.
The SATA SSD drive can do at most just under 550 MB/s.
On the other hand a good M.2 drive like the Samsung 970 Evo Plus can do 3500 MB/s in optimal conditions.
I will pick here a 1 TB drive, because NVMe drives get quite expensive the bigger they get. If we need more storage later we always have the option to add storage, either another m.2 drive, since our motherboard supports an extra drive, or via the SATA interface.
With an RTX 3090, you will want a computer monitor to match. My favorite brand is without a doubt Samsung.
I personally own the 34′ curved Samsung QLED monitor. It is great not only because it has a great screen. I use it not just for my own PC but also for my Macbook PRO. It can power it via the thunderbolt 3 port. There are few monitors better than this. It is not cheap… but then neither is the RTX 3090.