Best laptop for machine learning 2022
The use of machine learning and deep learning has led to a big revolution in the hardware required for these installations.
Another insight from the collector, or rather, the promised details about Threadripper: “In general, if the mantra “all memory channels must be filled” sits in you (and Threadripper has 4 of them), you can scroll down. I remind the rest that this processor of the old architecture has an interesting structure of 4 NUMA nodes, nodes with non-uniform memory access. They can work with a single channel, but then you will accumulate delays caused by this architecture.
Your own server for ML is like a machine in the 20th century: if you are seriously engaged in Data Science, sooner or later you will come to the conclusion that you need a single customized environment, confidence in resources that are independent of the rules of the employer and admins. Someone will say that everything can be done in the clouds, but constant access, long experiments 24/7, and even with data storage will cost a pretty penny. So what seems to be the Best laptop for machine learning 2022?
Let's see what we need:
- The datasets have to be somewhere. You need the ability to store and have quick access to large amounts of data.
- Video card. Resnets and Unets are waiting.
- Multi-core processor. Many people forget, but many operations in numpy, pandas, and even gradient boosting algorithms are perfectly parallelized for multithreading, but still do not work on video cards.
- RAM. Should include everything.
And the rest that should serve the parameters: a sufficient power supply, a case and a motherboard where all this happiness will fit, an operating system.
So, the first and most expensive is the video card. Let's face the obvious: Nvidia is now the absolute leader in performance and framework compatibility, and if you're writing neural networks, then you need cudnn and cuda. But good graphics cards cost a lot of money: if we want at least 11 Gb of memory and corresponding performance, we need to shell out $1000+ for a top model. Graphics cards are selling out and prices are going up. How to be? We must remember that right now we are living in a unique time: the cryptocurrency bubble is bursting at the seams, and a large number of video cards from miners are entering the market. I am the proud owner of a used 1080 Ti for 30 thousand rubles, and for almost a year it has never let me down, working 24/7. Take a computer with Windows (for some reason, most programs for testing video cards are specifically for this system), stock up on programs, test your workhorse up and down and feel free to take a graphics accelerator 1.5-2 times cheaper.
Continuing the idea of calculators, let's move on to the processor. Here I would like to say, by analogy with the previous one, that the absolute leader is Intel (especially since I once worked there). That's true...only if we're talking about single-threaded applications or an unlimited budget. However, we have neither one nor the other case, but we want to parallelize and leave money, here AMD Ryzen comes to us in general and their Threadripper line in particular. For $700-1000 you can buy a 24-32 nuclear chip, on which Catboost will fly, similar Intel parameters cost twice as much. Of course, there is a big “BUT”: Threadripper owes its performance to a specific design, and this will have to be reckoned with (more on that below) ...
Some lyrics from the builder: “I think after the recent Intel fiasco with the 10980XE, the question of which processor to choose for multi-threaded computing has a pretty clear answer. But… things can change.”
And then we turn to memory. Making a server with less than 32 gigabytes of RAM is strange (then it’s already easier to count on fitness bracelets) and it’s better to take memory with a high frequency (3200+, processors of the ZEN and ZEN 2 architectures love it). Of course, RAM is not the most complex component of the circuit, which means there are many manufacturers, but it is better to take proven ones (I took Corsair). Here you need to decide how much to take, and the number of channels. The simplest answer: more, so that each plate has 16 gigabytes. It seems that you can get 256 gigs of RAM in your PC. But not everything is so simple. If you take dual-channel memory, then twice as many active cores will access one amount of information in memory, which means that the access speed is reduced - here we must remember the need for fast memory access as a critical requirement. So we take a four-channel. On each plate we will have 8 gigabytes of memory.
Best Laptops for Machine Learning Under $1,000
The Acer Nitro 5 is a notable laptop for performing some heavy machine learning tasks. It comes with a wide range of features that make it the ideal tool for developers and engineers alike. You’ll benefit from features like the NVIDIA graphics card, fingerprint reader, connectivity options, upgradable Windows OS, multi-touchpad, backlit keyboard, and 11 hours of battery life.