You want to build a PC specifically for machine learning applications, but you don’t want to pay the exorbitant prices for specialized hardware? With a bit of luck, you’ll already have a good PC for use cases of this kind at home – because in terms of their hardware, they’re not dissimilar to gaming PCs.
What do I need a machine learning PC for anyway??
Roughly speaking, you feed your PC with very large data sets and train models to better handle certain tasks in the future. The computer thus develops algorithms on its own based on the data you gave it. Since we were just on the subject of gaming PCs, an example helps to understand the matter: Solutions are conceivable, for example, to significantly improve the image quality in games without compromising performance. NVIDIA’s in-house DLSS (Deep Learning Super Sampling) is an example of this method. On the other hand, these super PCs are also extremely useful for automated Machine Learning (AutoML), since a lot of experiments are performed. These can be executed in parallel, in order to reduce the runtimes of the programs significantly.
A subfield of these solutions are tasks like feature engineering: This very CPU intensive task optimizes the performance of existing algorithms. In other words, to cover all the fields, you need a lot of CPU and GPU power in one package.
The ideal PC for Machine Learning: buy or build it yourself?
The manufacturers of highly specialized machines for Machine Learning / Deep Learning of course pay a lot for their work. But in the end they only boil with water. This means: you don’t rely on magic hardware from a secret lab, but install RAM, GPU, CPU& Co. just like you would do it.
Purchased PCs of this type may use specially certified components (enterprise HDDs with particularly long warranty periods, ECC RAM and similar components), but the basic structure remains the same.
That means: The signs are good that you can assemble a completely sufficient PC for the most diverse machine learning models for little money. Let’s take a look at the important components in the overview:
RAM for ML computers
Here applies: Much helps much. The more RAM, the better. Since applications of this kind can never have enough RAM, you should not be petty here. You will shovel a lot of data between CPU and RAM and to the GPU if you want your algorithms to learn. Most data sets should be uncompressed to save time. 64 GB DDR4 RAM is the starting point, more can’t hurt.
CPU processor for ML computers
You have the choice between Intel and AMD, whereby AMD clearly has the lead here due to the current CPU landscape. Threadripper CPUs combine so many cores (and threads) in one package that Intel currently has nothing to compete with in this segment. A small calculation example: For a Threadripper 2920X with 12 cores and 24 threads, you’ll pay a bit less than 400 Euros. With Intel you get 8 cores and 16 threads for the same price. Although the single-core performance is higher there, this plays a subordinate role in machine learning applications. If you want to get your hands on one right now, there’s no way around AMD.
GPU graphics card for ML computers
The next big construction site is the GPU: A lot of VRAM and a high speed are important here. Keep in mind that VRAM works differently than RAM: If you don’t have enough VRAM, but your machine learning model demands it, it simply won’t start. A current GeForce Titan RTX with 24 GB VRAM will never have this problem, but it is also correspondingly expensive (just under 3 GB).000 Euro are due). You’ll find a better price-performance ratio in all GPUs with 8 GB VRAM or more – for example, an RTX 2080 Ti or even the older generation from GTX 1080 onwards.
Power supply for security
If most components are permanently running under high load, you also need a power supply that can withstand this load. Therefore, don’t save at the wrong end and invest in a power supply with a high efficiency level as well as enough power. If you play with the idea of installing several GPUs, the power consumption will increase rapidly. Of course we can’t give a general recommendation, because we don’t know what you will build into your PC in the end. When in doubt, buy an oversized power supply rather than an undersized one. In addition: Stay away from no-name products!
Water cooling to increase performance for Machine Learning
Waste heat has to be removed quickly – and your PC will produce a lot of this heat due to the high power consumption. Water cooling is a good investment here, because water conducts heat faster than air and on top of that lower temperatures are guaranteed. Even under high load components do not have to throttle or shut down. The better the cooling performance of your PC, the better the performance of your CPU and GPU chips. In approximately 30%-45% more achievement can be obtained thereby. In addition, the system then runs quieter than with pure air cooling, which can be important for your PC depending on the installation location.
You now have an overview of the most important components and what’s important – but does it really save that much money??
The comparison to NVIDIA’s DGX Station
With the introduction of the RTX 2000 series, NVIDIA has also introduced the DGX Station. This is a workstation designed for AI development, but can generally be used anywhere Data Science is done on a large scale. Equipped with 4 Tesla V100 GPUs (a Volta architecture-based GPU with 16 GB VRAM), a 20-core CPU and 128 GB RAM. For this hardware, NVIDIA is asking a hefty 49.000 US Dollar.
There is nothing to criticize about the performance of the DGX Station: Machine learning models that need to be trained get the job done about 50 times faster than on a single CPU. However, you can get fairly similar performance for a fraction of the price – and if you own a gaming PC, chances are you already have a powerful graphics card, plenty of RAM, and a fast CPU anyway.
Even offers from the cloud do not stand up to comparison: If you rent CPU/GPU computing capacity via Amazon (in the form of AWS) or via Microsoft Azure and use this for machine learning, you save around 90% of the costs when using a single GPU. What costs a good 3 euros per hour at AWS, costs only 20 cents at home. The advantage continues to grow in your direction when you switch on more GPUs.
Conclusion: Do-it-yourself is significantly cheaper – and almost as good
If you’re thinking of experimenting with machine learning, you can get comparable performance to the big cloud solutions or specialized hardware for just a few thousand dollars. For private customers, the investment in these professional solutions is practically completely unattractive – and there are hardly any differences to gaming PCs, with the exception of the huge RAM configuration.