Best laptop for Machine Learning

Table of Contents:

• MSI GS65
• HP Omen 15
• Acer Nitro 5
• ASUS ROG Zephyrus
• Apple MacBook Pro
• ASUS TUF
• Gigabyte Aero 15
• Dell G7

Many students and teachers study and work in public spaces. Finding the best laptop for machine learning that is also portable will be quite a difficult task, as ML will put a lot of demands on your laptop.

When it comes to AI and data science, you're only as good as your hardware. You will need something that can easily handle large data sets. A second key component is a modern GPU, so that your computer doesn't get too slow while performing huge calculations. (Trust us, nothing kills your creative flow like a great model that takes hours to complete build).

Laptops these days have different configurations to suit your purpose. Some of them come with high-end graphics for gaming. Some laptops have powerful processors and some have good memory. But what specifications do you need for your machine learning laptop? That is why in this review article we have compared all the necessary attributes that you need in your machine learning laptop.

MSI GS65 laptop for machine learning

Thin, light, fragile. The MSI GS65 was and remains one of the leaders among gaming laptops - which makes this model also a nice option for a machine learning engineer. Our test configuration is relatively simple and will cost $1849. It has a Core i7-8750H, 16 GB of memory, a GTX 1060 and a 256 GB NVMe drive. We take out the ruler and start the comparison!

GS65-8RF Stealth Thin, the predecessor of today's test device, was on our review recently - in May. It had a Core i7 processor, GTX 1070 Max-Q graphics and an NVMe drive. You can take this slim laptop with you and not be ashamed of its appearance, it provides an acceptable working time and has the same high-quality aluminum case as many more expensive models. We were pleasantly surprised by the RGB backlight, which is configurable individually, and the thin frames around the screen (which works with a refresh rate of 144 Hz and shows the correct colors). The performance turned out to be very good. At the same time, the price of that model, $ 2,100, gave us a reason to find fault with the quality of assembly and design, the quality of energy-saving capabilities, and even ill-conceived maintenance.

All current versions of the MSI GS65 laptop for machine learning use an Intel Core i7-8750H processor and either 16 GB or 32 GB of dual-channel DDR4-2400 memory. The GPU can be either GTX 1060 or GTX 1070 Max-Q. The size of solid-state storage varies from 256 GB to 1 TB, the connection type can be both SATA and NVMe, and there is even such an animal as a 1-terabyte "Super RAID 4" (and in fact these are two 512-gigabyte NVMe drives in a RAID-0 array).

Our current test sample is equipped a little worse than the past, it has Intel Core i7-8750H, GTX 1060 mobile graphics, 16 GB of DDR4 memory and a 256 GB NVMe drive.

As you probably should have expected, battery-powered performance drops significantly compared to mains power; running 3DMark 11 on standalone soldering gives out only 5758 points. But LatencyMon does not find any unnecessary delays.

When creating their new product, MSI tried to combine thoughtful design, compactness, lightness and power. The GS65 Stealth Thin boasts the world's first frameless IPS screen with 144Hz refresh rate. At the same time, the style of the machine remains quite neutral and will be suitable not only for gaming fans, but also for professionals of various professions, where power and mobility are important.

GPU

NVIDIA graphics cards up to GeForce GTX 1070 with 8 GB of GDDR5 memory are responsible for the graphics. The notebook has a pair of SO-DIMM slots for RAM with frequencies up to 2666 Hz. The storage system allows the installation of up to two high-speed M.2 SSD drives, combined in Super RAID 4. In general, is seems sufficient to use this laptop for machine learning purposes.

Display

You can connect several screens to this laptop, what expands the opportunities  (it is much more comfortable to code, build machine learnin models and test them on several screens). Here Matrix Display technology comes to the rescue, which allows you to create a configuration of 3 external monitors (plus the laptop display itself) with 4K/Ultra-HD (3840x2160) support.

All this "rich" stuffing fits into a 1.88 cm thick chassis, and the notebook weight is only 1.79 kg. A 15.6-inch IPS screen (144Hz) fits perfectly into a 14-inch laptop, thanks to a frameless display with a frame width of just 4.9 mm.

Battery

The battery capacity (82 Watt/h) in such a modest size does not suffer: the battery lasts for 8 hours of battery life in the so-called "office" mode. Cooler Boost Trinity system with 3 low-profile coolers and 4 heat pipes is responsible for cooling.

MSI GS65 also gets SteelSeries keyboard with customizable dot RGB backlighting. The SteelSeries Engine 3 utility not only lets you assign lighting and effects to each key individually, but also turns individual areas of the keyboard into interactive scales that display all sorts of useful data like character health, amount of ammo, mana, and so on.

Performance: new Intel Core i9 

Intel also got its share of presentations by unveiling the first Intel Core i9 processor for laptops (machine learning tasks require a lot of CPU capacity) and talking about the line of high-performance mobile products which will be based on it.

The main star of the series is the 8th generation Intel Core i9-8950HK, Intel's first mobile processor with 6 cores and support for 12 threads. The unlocked multiplier and support for the new Thermal Velocity Boost (TVB) feature allows the processor to boost up to 4.8GHz in turbo mode.

In addition, the company reminded about its fresh Intel 300 Series Chipset and the new extension of the Intel Core platform, which allows to combine the advantages of 8th generation Intel Core processors and Intel Optane memory: systems based on this platform will be marked with "plus" - Intel Core i5+, i7+ and i9+.

HP Omen 15 laptop for ML

The HP Omen line is aimed at the gaming audience, but its characteristics are such that for machine learning tasks this laptop is also a good fit.. The new Omen 15 is a very interesting device in terms of price-quality ratio.

Intel Core i5-10300H and Nvidia GeForce GTX 1660 Ti are a pretty decent basis for gaming in Full HD. The amount of RAM (16 GB) is sufficient for trouble-free operation of the system and applications, and a factory SSD of 512 GB is enough for the first time.

Intel Core i5-10300H contains 4 cores and supports Hyperthreading. The base frequency of this chip is 2.5 GHz and can reach 4.5 GHz according to Intel. The test laptop has a maximum frequency with a load on all cores of 4.2 GHz. However, even so, the i5-10300H loses to the AMD Ryzen 7 4800H in the MSI Bravo 15. In a single-threaded test, Intel still looks good, but in a multithreaded one, it lags very much.

The long-term performance of the Omen 15 processor looks good: the score in the Cinebench R15 multithreaded test (running in a circle) has not changed over time. After 50 iterations of the test, the average score was 897 points - Omen 15 slightly bypasses Nexoc GH15, based on the same Core i5.

Like most laptops for machine learning engineers and, in general, for developers, the keyboard is RGB backlit. It is divided into three zones across the width with a dedicated fourth zone of WASD gaming keys, you can assign different colors, blinking, color shimmers and other familiar effects, but only for the whole zones, not for individual keys. The proprietary utility Omen Gaming Hub sets the option of static backlight settings, and for advanced functionality you can run the utility Omen Light Studio. From the keyboard you can only enable the last used backlight mode and disable it, while the brightness has to be set in the mode settings.

Screen

The Omen 15-ek0039ur uses a 15.6-inch IPS display with a resolution of 1920×1080 pixels (MonInfo report). 

The grain of these defects is several times smaller than the size of subpixels (the scale of these two photos is about the same), so the focus on micro-defects and the "jumping" of focus over subpixels when you change the angle of view is weak, because of this there is no "crystal" effect.

The screen has good viewing angles without significant color shift even at large deviations of the view from perpendicular to the screen and without inverting shades. However, the black field becomes very bright and has a slight reddish tint when the screen is tilted diagonally.

Response time for black-white-black transition is 11.9ms (6.9ms on + 5.0ms off), the transition between grayscale halves in total (from hue to hue and back) takes an average of 14.4ms. The matrix is fast, but there is no acceleration in an obvious form - there are no brightness bursts on the transition fronts. 

Battery life

The laptop's battery capacity is 71Wh. Screen brightness is set at 100 cd/m² (which in this case corresponds to about 56%) when testing, so laptops with relatively dim screens don't benefit. The laptop doesn't allow you to set the battery near full discharge level below 7%, so we tested the battery discharge to this value (and the charging process from the same).

The first battery life tests gave low figures, which were in direct conflict with the 12 hours of battery life declared on the manufacturer's website. We tried to honestly improve the results, first switching on hybrid graphics mode (i.e. in our scenarios discrete graphics card was switched off), then changing the CPU mode, and finally switching off the keyboard backlight (for word processing scenario this is wrong, but really useful for model training, for example). I couldn't get up to 12 hours but the notebook somehow made it through 7 hours of work with text. When watching movies it was enough for 5 hours. In general, the level of battery life of the laptop can be called average.

This machine learning laptop does not have any charging settings. It charges in a little less than 2 hours: at the beginning the process is uniform and takes one hour to reach 81%, then the process slows down a lot and it takes another hour or so to get the battery back to 100%. All in all, this is the typical picture of a fast charge. The information LED on the left side of the case shines orange when charging, switches to white only when the charge reaches 100%, blinks white when discharging below 10%.

Performance under load and heating

Traditionally, the laptop has two coolers with multiple heat pipes. In this case there are three such tubes, two of them are "common", that is, they cool the CPU and GPU simultaneously and reach both radiators. Thus, both coolers are at least passively always involved in cooling, but we do not know the exact algorithm of their work: there is no fan speed monitoring. We can only qualitatively estimate (by touch) that when one component (CPU or graphics card) is under load, both fans rotate. Cold air is sucked in from the side of the case bottom, there are enough ventilation holes. The fans blow heated air to the back and to the right, the latter is not the best option, as it heats the hand on the mouse.

The Omen Gaming Hub utility offers two performance profiles: Performance and Balanced. You can also separately adjust coolers, but without any specific fan speed numbers - only selecting a range with a certain name is available. Naturally, we tested both profiles, and left the control of cooler operation to automatics. Note, by the way, that the laptop does not have the usual profile for maximum power saving or maximum silence - HP Omen seems to be exclusively a gamer's product.

In Maximum Performance mode, when the laptop is busy with machine learning objectives and the processor is under load, there is a traditional for Intel products spike in consumption when Turbo Boost is on - up to 105W. The temperature rises up to 90 degrees very quickly, but no trolling is observed. About a minute later, the fans increase the speed significantly, the noise increases significantly, and then Turbo Boost shuts down, so that the processor goes to a stable 90W consumption (to put it mildly, a little above the normal TDP level of 45W) at 3.8GHz cores and with a temperature slightly below the same 90 degrees. Coolers can slow down and overclock again, which affects CPU temperatures.

When the CPU and graphics card are under simultaneous load for the first time there is a trolling - up to half of the cores at the same time. After the initial surge (up to the same 105W) the processor tries to roll back to 90W, but it constantly overheats, so it has to lower the consumption bar more and more, and in the end it stops balancing on the edge of trotling at 3.5-3.55GHz cores and consuming about 75-76W. Its temperature is literally a couple of degrees off, a couple of cores are trolling, then they stop. We believe that in normal life, when the load on the processor will not be synthetic stressful, but real, overheating in such mode will not occur. The video card slightly decreases performance in relation to the scenario without simultaneous CPU load, but they remain stable: 90 W consumption and GPU frequency 1200 MHz. Noise from the coolers is maximal and very high.

In Balanced performance mode the surge (up to 105W) when the CPU load is applied turns out to be quite short, for 15 seconds, after which the processor steadily consumes 55W (still above TDP), it is far from overheating and trolling. Cooler noise is quite acceptable.

Noise level

We measure the noise level in a special soundproofed and half-silenced chamber. The microphone is placed in relation to the laptop in such a way as to imitate the typical position of the user's head: the screen is tilted backwards by 45 degrees (or maximum if the screen is not tilted back by 45 degrees), the microphone axis coincides with the normal coming from the screen center, the microphone front end is at 50 cm distance from the screen plane, the microphone is directed to the screen. The load is created using the program powerMax, the brightness of the screen is set to maximum, the temperature in the room is kept at 24 degrees, but the laptop specially not blowing, so in the vicinity of the air temperature may be higher.

If a notebook does not load at all, even in Balanced mode, its cooling system can not work long in passive mode - after a while the fans turn on and do not turn off. Under heavy load the noise from the cooling system, as well as the performance, depends on the selected profile. Forcing the fans to maximum speed results in a significant increase in noise level and a slight increase in performance. The character of the noise under average load is smooth and does not cause irritation, but at high fan speed there is a low-frequency hum (corresponding to the frequency of beating of two fans), which is already irritating. In general, the laptop, of course, a lot of noise under heavy load, including machine learning processes. Unfortunately, this is typical for all models with a discrete graphics card, so it is better to play on gaming laptops with headphones.

Performance

The HP Omen 15-ek0039ur has an Intel Core i7-10750H processor from the last, 10th generation (Comet Lake). It's a 6-core (12-core) processor with up to 5.0 GHz and a TDP of 45 watts. As testing under load has shown, the processor in this laptop is always running at a higher consumption: 55W in balanced mode, and 90W in performance mode. The fact of exceeding the TDP of the processor in laptops, to put it mildly, is not a rarity (after all, this value says exactly about the recommended cooling, and it is different in each laptop), but has it long been the norm to exceed the TDP twice? Nevertheless, there are no claims for processor stability even at 90-watt consumption, no overheating and no trolling, so we have a very productive processor.

In tests that have little dependence on the storage the HP Omen turned out to be faster than the recently tested MSI GP66 Leopard 10UG with the same processor (in that case, as we remember, the processor was regularly trolling). In comparison with the selected rivals the general conclusions are the same: the result is far from being a record, even Intel processors are faster, not to mention the older modern Ryzen 4000 and 5000, but for the Intel platform it is close to the current (before the release of productive Tiger Lake) maximum. All in all, it's a high level, enough for any reasonable laptop application, even more so for gaming.

SSD performance in real-world tasks is also not the maximum, but it is very high, the bottleneck of such a drive will definitely not become a bottleneck.

On the whole, we liked the HP Omen 15. It is definitely a good choice among the best laptops for machine learning, allows to play in Full HD with high quality graphics, displaying the picture on the fast screen with a refresh rate of 144 Hz, although the video card here is not top-end by today's standards (but in the last generation it was one of the top solutions). At the same time, the laptop is quite universal: it has a fast processor, a decent amount of expandable memory, you can install a second SSD-drive, enough interface ports, there is a card drive. The keyboard will please you with a customizable RGB backlight and an excellent, unique layout for text editing.

Acer Nitro 5 laptop for AI

The Nitro 5 series is very popular laptop for machine learning among AI professionals.. There is nothing surprising in this, because we are dealing with inexpensive solutions (for example, against the background of the Predator Helios series), which nevertheless have decent performance regarding ML tasks. So, the configuration of the Acer Nitro 5 AN515-54-56MH laptop for machine learning includes a 4-core Core i5-9300H chip and GeForce GTX 1660 Ti mobile graphics - listing these characteristics is enough to say: this is a fairly productive gaming PC.

I will immediately draw your attention to the fact that the updated Nitro 5 series includes not only components from Intel and NVIDIA. On sale you will also find configurations based on AMD chips. This laptop can use Ryzen 3000 processors and Radeon RX 500 graphics. All the main characteristics are listed in the table below.

Keyboard

Importantly for machine learning, the laptop uses a full-size island-type keyboard with a separate number block and red backlight. The top row of keys is equipped with additional functionality, which is activated in combination with the [Fn] button.

The keys have flat caps and the following sizes: main - 15 x 15 mm, functional - 10 x 12 mm, arrows - 15 x 15 mm, number pad - 12 × 15 mm. The keyboard is very pleasant and comfortable. The keys are easy to press and have a short stroke. At first this led to accidental presses, but we adapted to this very quickly.

The touchpad is quite standard in its size (112 × 79 mm). It also has a backlit outline of the working surface. To the touch, the coating is slightly rough, and the finger slides with slight jerks, although the cursor positioning is clear and smooth.

Display

More expensive modifications may have a panel with a refresh rate of 144 Hz, but in our case we use the IPS panel BOE NV156FHM-N48 with a resolution of 1920 x 1080, glossy coating and refresh rate of 60 Hz. The thickness of the bezels is as follows: the side ones are 8 mm, the top one is 15 mm, and the bottom one is 27 mm.

The screen pleased with good brightness, contrast and viewing angles, but the amount of color space sRGB clearly pumped up. In general, it is quite good laptop for machine learning (not that much for a designer or artist, on the contrary).

Performance

The laptop modification for machine learning  - Acer Nitro 5 AN515-44 with the index NH.Q9GEU.00J - is equipped with a 6-core 12-core AMD Ryzen 5 4600H processor. Its base frequency is 3.0 GHz and its dynamic frequency reaches 4.0 GHz. TDP is 45W.

The AMD Radeon Vega 6 video core integrated into the processor includes 384 stream processors and runs at 1500 MHz. Memory for the needs of the graphics subsystem is allocated from the RAM. As it works in single-channel mode in the laptop, we get a 64-bit bus with up to 25.6 GB/s of bandwidth. If you install a second memory module, you get a 128-bit bus with a bandwidth of 51.2 GB/s in dual-channel mode.

GPU

More interesting for machine learning engineer, however, is the NVIDIA GeForce GTX 1650 discrete graphics card. At its core is the NVIDIA TU117 GPU, which has 896 CUDA cores, 56 texture and 32 bitmap blocks. The frequency formula is 1380/1515 MHz for base and dynamic values respectively.

The video memory is made up by means of GDDR6 chips from Micron company with total amount of 4 GB. Their effective frequency is 12000 MHz. The bandwidth reaches 192 GB/s with a 128-bit bus.

The RAM subsystem is represented by a single 8GB card, which operates in single-channel mode at 3200 MHz.

The disk subsystem is implemented based on a 512 GB M.2 SSD WD PC SN530 (SDBPNPZ-512G) with PCIe 3.0 interface. The sequential data reading speed reaches 2481 MB/s and writing speed is 1813 MB/s, which is a decent performance.

There is absolutely no question about the performance of the AMD Ryzen 5 4600H + GeForce GTX 1650 tandem. You can get an additional bonus after installing another memory module and the implementation of dual channel mode. In general, the figures are a little lower than the combination of Intel Core i7-9750H + GeForce RTX 2060 (Acer Predator Helios 300).

In the default operating mode, the frequency of the graphics chip varies from 1650 to 1785 MHz. At the same time, in the game — that is, in an application where the CPU and GPU are simultaneously involved — the cooling system works noticeably louder: from a distance of 30 cm, the equipment recorded 41 dBA. Noisy, but quite tolerable, in my opinion. At the same time, the GPU heats up to 83 degrees Celsius, but the temperature of the hottest core of the CPU reaches 96 degrees Celsius - the design features of the Nitro 5 cooler, which I talked about in the first part of the article, affect. As a result, the laptop heats up quite a lot.

Cooling system, noise level and temperature mode

Active cooling system is responsible for the removal of heat from the internal components. It includes two copper bases for the CPU and GPU, three heat pipes and a pair of fans combined with small radiators. The air temperature in the test lab was 25°C at the time of measurement.

At idle time, the temperature inside the Acer Nitro 5 AN515-44 was 44°C for the CPU and 30°C for the GPU.

During the stress test, the CPU temperature rose to 76°C and the GPU to 57°C. There was no processor trattling.

In general, the cooling system is at a very decent level in terms of performance, and does not stand out with excessive noise. Of course, it cannot be called quiet either, but everything is learned by comparison. The case heating is OK, because the temperature of the working area did not rise above 38°C.

Power supply, battery and battery life

The device comes with a 135 watt PA-1131-26 power adapter. The input signal parameters are 100 - 240 V at 1.9 A and 50 - 60 Hz. The output is 19.5V at 6.92A.

This machine learning laptop is equipped with a non-removable 4-cell lithium-ion battery with a capacity of 3733 mAh. In reality, you can expect 3-4 hours of operation in a mixed usage scenario. It is also worth making a correction that autonomy testing was done at maximum fan speed of the cooling system. If the cooler will work in automatic mode, the battery life should be longer.

Acer Nitro 5 AN515-44 in modification NH.Q9GEU.00J is an interesting solution in its segment. At the cost of about 24,999 UAH ($885) it offers not only an excellent set of characteristics and good performance, but also a good head start on upgrades.

The NH.Q9GEU.00J version has only one 8 GB RAM and a 512 GB SSD. In the future, you will be able to buy a second memory slab and two more drives, because there is not only a free M.2, but also a slot for a 2.5-inch drive.

Machine learning capabilities of this laptop are entrusted to a bunch of 6-core 12-thread AMD Ryzen 5 4600H processor and NVIDIA GeForce GTX 1650 graphics card. This is enough to run online projects and actual blockbusters. The list of advantages can be continued by an efficient and relatively quiet cooling system, a high-capacity battery and the manufacturer's attention to the quality of the camera and microphone.

ASUS ROG Zephyrus laptop for machine learning

The Asus ROG Zephyrus S17 laptop for machine learning, as you might guess from the name, has 17-inch screens, so these models can't be compact or light. But not everyone needs portability — after all, gaming laptops generally assume primarily stationary use. But the performance of the Zephyrus S17 GX703HS model we are testing is excellent, as a highlight we will allocate a RAID array of three (!) terabyte SSDs. And then there's Intel Core i9-11900H and Nvidia GeForce RTX 3080 Laptop, 32 GB of memory and an interesting keyboard and wheel.

Keyboard

The keyboard here does not take up much space, there are large fields around it, but there is logic in this: for a gaming notebook it is more important that the user can comfortably place his hands and press the right button in the right fraction of a second. In general, the layout is progressive, there are positive changes relative to what we usually see in laptops. The keys of the top row of the keyboard are reduced in height, but divided into standard blocks of four F-keys with significant gaps between them. 

"Arrow keys" have the same (although reduced) size and are more or less separated, even beyond the standard bottom row (at the same time, the "thick" space bar of complex shape goes there as well). The keys for text editing (Home/End, PgUp/PgDn, Ins/Del) are also slightly separated. There are two Fn buttons at different ends of the keyboard, so there shouldn't be any problems with handy chords. 

There are also 4 additional keys above the standard keyboard block - for operative work with volume and microphones and for launching the proprietary Armoury Crate utility. The power button is also placed outside of the keyboard, so you can't accidentally press it. By default, the upper row of keys performs the functions of F1-F12, and to control the notebook settings (media player buttons, change the brightness of the screen, switch the cooling system profile, turn off the touchpad and wireless networks...) these buttons have to be pressed with Fn. We have not found the switch of this mode in the laptop.

The keyboard has a membrane mechanism and island key arrangement, the keys are moderately large, comfortable: the distance between the centers of the keys in one line is 19 mm (slightly larger than usual), and between their edges - 3 mm. The keys are expectedly very quiet, but also with a rather elastic pressure, so that the feedback here is felt better than in conventional membrane keyboards. And this is despite the fact that the stroke of the keys is very small - 1.2 mm. We should also mention that the handling of presses is independent (n-key rollover), that is, no matter how many keys you press simultaneously in the heat of battle, the game reacts to all of them.

There is a three-level brightness white backlight (the fourth state - off). The characters on the keys are highlighted and slightly - their contours, but at a deviation from the keyboard under each key you can see the backlighting area. However, it glows softly and does not irritate your eyes at all.

The screen

The Asus ROG Zephyrus G15 GA502IU uses a 15.6-inch Sharp IPS matrix with 1920×1080 resolution.

The outer surface of the matrix is black hard and semi-matte (the specularity is well defined). There are no special anti-reflective coatings or filter, and there is no air gap. When powered from the mains or battery and with manual brightness control (there is no automatic adjustment by the light sensor), its maximum value was 270 cd/m² (in the center of the screen against the white background). 

Note that by default there is an automatic brightness adjustment of the backlight depending on the brightness of the image (brightness is reduced for dark scenes), but this function can be disabled in the settings of the graphics core. The maximum brightness is not very high. However, if you avoid direct sunlight, even this value allows you to somehow use the laptop outside even on a summer sunny day.

If the brightness setting is 0%, the brightness goes down to 13 cd/m². In complete darkness, its screen brightness can be lowered to a comfortable level.

There is no flicker (neither visible nor detectable in the stroboscopic test) at any brightness level. If we take a very strict approach, the brightness-time relationship at low brightness reveals the presence of modulation, but its nature (frequency of the order of 24 kHz and a small amplitude relative to the maximum brightness) is such that flicker is never and under no circumstances detected and can not in any way affect the user's vision.

Battery operation

The battery capacity of the laptop is 76Wh. Screen brightness is set to 100 cd/m² (which in this case corresponds to about 48% brightness) when testing, so laptops with relatively dim screens don't benefit.

In general, the battery life of the laptop during machine learning tasks is quite decent, although far from a record. The battery has an impressive volume, so that there is an attempt to create a hybrid of a powerful gaming laptop and portable. Of course, we're not seriously discussing battery life in games or under any other heavy load. But if you want to combine work on the run and gaming in the comfort of your home, this laptop is quite suitable.

The full battery charging time from the regular adapter is about 2 hours. The MyAsus proprietary utility allows you to enable a battery life extension mode, selecting how much to undercharge the battery, according to the typical usage profile of the laptop. The corresponding LED on the case lights red when charging (up to 95%) and white when running, blinking red when below 10% discharge.

Asus offers a lighter, more compact 65-watt option for quick charging when you're on the go. This adapter has a USB Type-C output jack, so you can use it to charge smartphones and any other mobile equipment as well. Conversely, if you already have one (with a modern smartphone, for example), you can use it to charge the Asus ROG Zephyrus G15 without any problems. Also, of course, supported charging notebook from external batteries (power banks).

Work under load and heat

The Asus ROG Zephyrus G15 laptop for machine learning has two hot chips: a 60-watt discrete video gas pedal Nvidia GeForce GTX 1660 Ti (Max-Q) and a 35-watt AMD Ryzen 7 4800HS processor. The cooling system used here is traditional for such cases: small pads on the CPU and GPU and several heat pipes to the radiators at the back and right side of the case, through which the air is blown by two fans. Cold air is mostly sucked in through the holes on the bottom of the case, and hot air is blown backwards (at that it does not disturb anybody and does not even heat up the lid) and to the right (a right handed person lying on the mouse can get a little warm). Note that the heat pipes intertwine the two coolers in such a way that even if the load is only on the CPU (or only on the video gas pedal) both fans still work, and together they cool harder.

The laptop uses very hot components, and the thickness of the case, although not small, is standard, so here we see the usual cooling system that uses the internal space to the maximum: 6 heat pipes that are pressed not only to the CPU and GPU, but also to the elements of the power converters on the board and video memory. There are formally two coolers (with 2 fans and 4 radiators in total), but they are so intertwined that in fact it is a single cooler. And the fans, when loading any of the components (processor or video card), work synchronously. The Thermal Grizzly liquid metal alloy is used as a thermal interface on the processor.

To start with, let's note that when the ambient temperature is low enough, the laptop's fans can stop at idle with the Silent profile. With the Performance and Turbo profiles they are always running: at 2800/2600 and 3600/3400 rpm respectively. CPU and GPU power consumption at idle is 3 and 4W.

Now about what happens under load. Here the Silent profile is fundamentally different from the other two. It has significantly lower maximum fan speeds (up to 4500/4300 rpm), and this leads to the fact that not only the CPU, but also the video gas pedal at some point limits consumption at a reduced level. Reduced CPU consumption (25W) may still be considered normal, but halved (from 60 to 30W) video card consumption indicates that the Silent profile is not suitable for gaming. Probably, it will be relevant when the system performance is not critical - or when silence is more important.

In Performance and Turbo profiles, the notebook cooling system behaves roughly the same, but in Turbo the cooler frequency is always higher (including higher maximum frequency): up to 5900/5900 and up to 6700/6100 rpm respectively (the latter being the physical maximum of the coolers). Admittedly, aside from lower component temperatures, in the long run this provides a clear Turbo benefit, only when under maximum simultaneous CPU and graphics card load with the Performance profile does the CPU still reduce consumption below its normal 35W (to 30W, core frequency drops to 2.65GHz), while with the Turbo profile consumes 35W until the winning end (and the core frequency remains at 2.80GHz). Of course, the improved cooling also affects how quickly the processor "drops" to normal consumption from short-term increased consumption (it's 42W in both profiles), so in real-life tasks where performance is needed in short bursts, the laptop can be significantly faster with the Turbo profile. The graphics card in both profiles performs the same, its consumption is always close to the normal 60W.

When applying the maximum load to the CPU, a typical pattern is observed for Intel processors: at first, spurred by the Turbo Boost mode, the processor operates in excess of all reasonable limits, including its own TDP (which, according to official data, is only 45 watts). The frequency of cores under load exceeds 4 GHz, consumption is in the region of 105 watts, the temperature rises to almost 100 degrees and at least one core begins to trot. However, this period does not last long, just a couple of minutes. Then the consumption decreases to still very high 90 watts, the temperature drops below 90 degrees, the frequency stabilizes at 3.75 GHz, the trolling disappears. This is a stable long-term regime. But the operation of the cooling system in this case is far from ideal, the fans are constantly lowering, then raising the rotation speed (approximately in the range of 2700-4600 rpm), aggravating the high noise level by also changing its volume and tonality.

On the whole, the notebook coolers cope with the task of cooling the components, as such overheating and trolling are not observed in any load scenario, but limitations on consumption and performance under heavy load are still apparent in all profiles, except Turbo. So for the purposes of our testing we will limit ourselves to this profile, and the necessity of the other two is quite clear: you do not always need maximum speed, sometimes silence is more important.

Noise level

We measure the noise level in a special soundproof and half-silenced chamber. The microphone is placed in relation to the laptop in such a way as to imitate the typical position of a user's head: the screen is tilted backwards by 45 degrees (or maximum if the screen is not tilted back by 45 degrees), the microphone axis coincides with the normal coming from the screen center, the microphone front end is at 50 cm distance from the screen plane, the microphone is directed to the screen. The load is created using the machine learning model training tasks, the brightness of the screen is set to maximum, the temperature in the room is kept at 24 degrees, but the laptop specially not blown, so in the vicinity of the air temperature may be higher. 

Apple MacBook Pro laptop for ML

The undoubted leader and a true favorite of developers. The power of this laptop for machine learning perfectly fits the concept of programming not only in 2022, but also in the coming years. The price bites, but the MacBook Pro 16″ is totally worth the money. Choose the powerful M1 Pro chip or the even more powerful M1 Max to boost your workflow performance to a professional level.

The new MacBooks with Liquid Retina XDR display are powered by M1 Pro and M1 Max processors — the most productive chips in Apple's history.

The device can be the best laptop for machine learning and complex mathematical calculations. According to the developers, "MacBook Pro cope with tasks that were previously simply unthinkable for laptops."

Also, the new MacBook Pro received a 1080-pixel FaceTime HD camera, an improved audio system with powerful bass and a large set of connectors.

Up to two Pro Display XDR monitors can be connected to models based on the M1 Pro chip. And models with the M1 Max chip support simultaneous operation of up to three Pro Display XDR and a 4K display.

The updated design made it possible to improve the cooling system: the new design allows 50% more air to pass through, so the fan does not even turn on when performing everyday tasks.

The keyboard has also changed. Laptops received Magic Keyboard keys with a Force Touch trackpad: Apple abandoned the screen at the top of the layout and returned to physical buttons.

Sound with spatial audio technology is output to six high—definition speakers - two high-frequency and four low-frequency with resonance correction. This provides 80% more powerful bass. The models are also equipped with three professional studio microphones with noise reduction. The developers assure that this is the best audio system that Apple devices have ever received.

The performance of M1 Pro is up to 70% higher than that of M1: in the maximum configuration, the new chip is equipped with a powerful central processor - up to 10 cores - and a 16-core GPU. Laptops based on the M1 Max chip will be even more powerful: instead of 16 cores in the GPU, they got 32.

M1 Pro and Max reach peak performance values by consuming less power than similar chips.

The bandwidth of M1 Pro is 200 GB/s. The M1 Max has 400 Gb/s. This is three and six times more than the M1. The chips are equipped with fast combined memory - up to 32 and 64 GB, respectively.

Thanks to the built-in 16-core Neural Engine system, both new chips provide high-speed processing of machine learning processes and more efficient camera operation.

ASUS TUF laptop for machine learning

For the demanding, Asus produces thin and sexy ROG Strix laptop for machine learning and ROG Zephyrus laptops, but thrifty gamers are looking towards the TUF Gaming family, which is distinguished by an extremely attractive combination of power, availability and reliability. With an emphasis on reliability, as befits the products of the TUF family: the filling of laptops is hidden in metal-plastic cases that are certified by the US Army (MIL-STD-810G standard) and can be used to work in adverse conditions.

Screen

The Asus TUF Gaming F15 laptop for machine learning uses a 15.6-inch IPS matrix with a resolution of 2560×1440 (edid-decode report).

The outer surface of the matrix is black hard and semi-matte (specularity is pronounced). There are no special anti-reflective coatings or filter, and there is no air gap. When powered from the mains or battery and with manual brightness control (there is no automatic adjustment by the light sensor), its maximum value was 265 cd/m² (in the center of the screen against the white background). The maximum brightness is not high. However, if you avoid direct sunlight, even this value allows you to somehow use the laptop on the street, even on a summer sunny day.

Battery life

The battery capacity of the laptop is 90Wh.  Like many modern gaming laptops, the battery capacity here is very impressive. Well, perhaps this will contribute to the occasional use of the Asus TUF Gaming F15 (2022) as a portable laptop for machine learning, that is, will save the purchase of a second laptop specifically for travel. 

In general, there are no obvious contraindications to this, except for the gaming video card and clearly gaming positioning: the device is not so heavy and bulky, except that the brightness of the screen is not enough to work in the sun. But in any case it should be understood that battery life in typical "standalone" scenarios (working with text and watching video) here is not very impressive compared to office laptops without a discrete graphics card: the Asus TUF Gaming F15 (2022) gets 6 and 4 hours, and those could easily be, say, 10 and 6 hours with the same load. And about the gaming use of the laptop away from the mains we can not talk at all: you can count on an hour maximum, and that with reduced graphics card performance. By the way, the table also provides a clear demonstration of why you should not always use a discrete graphics card for work, even if you are happy with the Mux-switch: battery life even under the lightest load is reduced by half.

Performance under load and heating

Cooling system in Asus TUF Gaming F15 (2022) is serious, two hot chips (CPU and GPU), as well as components of their power converters are covered by five heat pipes, distributing heat to four radiators with two fans. When one of the chips is under load, both fans always work (in fact, both always work at idle, just very quietly), so we can say that the cooler here is a single one. The fans blow heated air to the back and to the right/left, on the hand with a mouse blows hot (when the laptop is working under load). The case in the back also heats up a lot, but the screen in this case does not suffer.

At idle in the maximum performance profile Turbo, the laptop is very quiet, because both fans are engaged, just at low speed. This distinguishes this model from those in which, for unknown reasons, one cooler is responsible only for the video card and does not turn on until you load it. It does not matter if the second cooler overheats or howls. In this case we don't even have to advise to switch to economy profile after quitting the game or after finishing work - the noise will be the same, only periodically fans will stop, and just regular switching from complete silence to very quiet work and back is more noticeable by ear.

At maximum processor load in the Turbo profile, its P-core frequency goes up to 4.1 GHz (their nameplate maximum frequency is 4.7 GHz, but this is clearly for when only one or two cores are loaded), E-core frequency to 3.15 GHz (regular - to 3.50 GHz), and consumption - exactly to the claimed 115W. During this surge the processor temperature can approach 100 degrees, one core can manage to overheat and start trotting. But the period of increased consumption lasts about half a minute, then the frequency of cores (of both types) decreases significantly, and the consumption drops to the normal 45W. Tit for tit!

Noise level

Initially, Asus set the bar at the level of a psychologically comfortable figure for buyers of $1000. Because of this, laptops for machine learning had to cut off excess fat like USB C with Power Delivery, fingerprint scanner, fast charging or RGB backlight. Over time, the TUF family has grown, and the boundaries have blurred, but the principle of "nothing superfluous" has not gone away. Unlike the flagship models of the ROG series, TUF laptops have only the most necessary for a gamer: productive hardware, a display with a high refresh rate, a convenient keyboard and an efficient cooling system.

We measure the noise level in a special soundproofed and half-silenced chamber. The microphone is placed in relation to the laptop in such a way as to imitate the typical position of the user's head: the screen is tilted back 45 degrees (or maximum if the screen is not tilted back 45 degrees), the microphone axis coincides with the normal coming from the screen center, the microphone front end is at 50 cm distance from the screen plane, the microphone is directed to the screen. The load is created with powerMax program, the screen brightness is set to maximum, the temperature in the room is kept at 24 degrees, but the laptop is not specially blown, so the air temperature in the vicinity of it can be higher. 

Gigabyte Aero 15 laptop for machine learning

Unfortunately, due to the well-known insanity, there are practically no new GeForce RTX 30 series video cards in retail. And although Nvidia seems to be starting to struggle with miners (let's see how things go with the GeForce RTX 3060), right now it's not easier for ordinary buyers. For them, almost the only option to purchase a card like the GeForce RTX 3080 is to buy a laptop with it. However, it will be a mobile version of the accelerator, by definition slower. And the rest of the laptop components are not exactly given away for free. But if you were going to buy a gaming laptop anyway, at least you have the opportunity.

The last letter of the full name of the model (C) tells us that this is the third generation of Gigabyte Aero laptops. The first generation (A) was based on Intel Core processors of the 9th generation, the second (B) - on Intel Core of the 10th generation, in both cases Nvidia GeForce RTX 20 video cards were used. The current generation still has the same Intel processors, but the graphics cards have been updated. As a result, the three models of the current YC, XC and KC family differ from each other in just the video card — they use GeForce RTX 3080 Laptop, GeForce RTX 3070 Laptop and GeForce RTX 3060 Laptop (for machine learning it's a good one!), respectively.

Keyboard

The keyboard has a membrane mechanism and the island key arrangement of the standard size (15×15 mm), the distance between the centers of the keys in a row is 19 mm (slightly larger than usual), and between their edges - 4 mm. The keys are very quiet. The stroke of the keys is standard - 1.5 mm. We should also mention that the handling of presses is independent (n-key rollover), that is, no matter how many buttons you press simultaneously in the heat of battle, the game reacts to all of them.

There are two-level brightness RGB-lighting (the third state - off), individual for each key (per key RGB). The characters on the keys themselves and slightly, softly, their contours are highlighted. The characters of the Russian keyboard layout are not illuminated, it may be a problem for those who can not type blindly and work in a darkened room. When you press Fn, the keys that perform some other function in combination with Fn are highlighted (bright white, ignoring the general settings). However, the designations of these other functions (lettering or icons) are also placed on the opaque part of the keys; they themselves are not illuminated, which would be more logical. (For example, when you press Fn on the F11 key, the airplane could light up, and you would not have to remember which of the function buttons activates the "in airplane" mode).

Screen

Let us note that on the screens with a wide color gamut without the corresponding correction the colors of ordinary images optimized for sRGB devices look unnaturally saturated. By the way, as a rule, in developed OS, in Windows in particular, and/or in more or less advanced software for working with images, the necessary color correction is achieved by using a color management system (a color profile for the notebook screen is already preinstalled in the system, and the display itself is calibrated in factory conditions). Therefore, the wide color gamut is not a disadvantage in this case. Some difficulties with obtaining the correct colors may arise in games and when watching movies, but even this, if desired, can be solved.

The Gigabyte Aero 15 OLED XC laptop for machine learning screen has a high enough maximum brightness (363 cd/m² in SDR mode) that the device can be used on a bright day outside, shielded from direct sunlight. In total darkness, brightness can be reduced to a comfortable level (up to 4.5 cd/m²). The undeniable advantages of OLED-screen can include a true black color (if the screen is not reflected), a noticeably smaller than the LCD, the drop in brightness when viewed from an angle and excellent support for HDR (high peak brightness, infinite contrast, wide color gamut, increased number of shades). The disadvantages include zone flickering at low brightness. In general, the quality of the screen is very high, but in case of professional use you should take into account a slight clutter in the shadows, as well as barely noticeable, but still detectable static noise.

Battery operation

The battery capacity of the laptop is 99Wh. To give an idea of how these figures compare to actual battery life, we test using our methodology. The brightness of the screen is set to 100 cd/m² (which in this case corresponds to about 53% brightness) when testing, so laptops with relatively dim screens don't get an advantage.

The components in the Gigabyte Aero 15 OLED XC are far from the most economical, this is not an ultra-mobile processor without a discrete graphics card, and system consumption even at idle can be 40 watts. Usually, we usually choose the performance profile for the battery life test, so as not to interfere with the laptop to perform its tasks (another thing is that in our standard scenarios the load on the processor is quite small). However, the Aero 15 OLED XC in this mode while typing drained only 2 hours.

Working under load and heating

The Gigabyte Aero 15 OLED XC laptop for machine learning has a pretty powerful cooling system with 2 fans and 4 heat pipes. However, only one tube goes from the CPU and GPU heatsinks to "their" fan, the remaining two tubes are "common", so there is actually a single cooler rather than two separate ones. This is confirmed by the cooler operation scheme: when any of the components is loaded, both fans turn on and run at the same speed. 

Cold air is sucked in from the bottom, where almost half of the bottom of the laptop is covered only by the grid. Hot air is blown to the back (mostly) and to the right/left at the back of the case (to a lesser extent). The heating of the case near the vents and in general in the back, of course, is felt, but in normal life is not particularly disturbing. The maximum speed of the fans is 5300 rpm, and they produce a very high level of noise. This emergency mode can be promptly turned on and off by pressing Fn+Esc without the use of a proprietary utility.

At idle, the laptop fans run at about 2000 rpm in Normal and Gaming modes, and in Quiet mode they stop, with no concerns about component heating. The AI control unfortunately doesn't stop the fans, so if you want complete silence, you'll have to manually switch to Quiet (or set your own mode). Even at 2000 RPM, though, the laptop is very quiet.

Noise level

Gigabyte Aero 15 OLED XC has a fairly powerful cooling system with 2 fans and 4 heat pipes. However, only one tube departs from the radiators on the CPU and GPU to the "own" fan, the two remaining tubes are "common", so there is actually a single cooler, not two separate ones. This is also confirmed by the operation scheme of the coolers: when loaded on any of the components, both fans turn on and operate at the same speed. Cold air is sucked in from below, where almost half of the bottom of the laptop is covered only with a mesh. The hot air is blown backwards (mostly) and to the right/left in the rear of the case (to a lesser extent). The heating of the case near the vents and in general in the back, of course, is felt, but in ordinary life it does not interfere much. The maximum speed of the fans is 5300 rpm, while they produce a very high noise level. This emergency mode can be quickly turned on and off by using the Fn+Esc keyboard shortcut without using the proprietary utility.

We measure the noise level in a special sound-insulated and half-silenced chamber. In this case the microphone of the noise meter is placed relative to the laptop so as to imitate the typical position of the user's head: the screen is tilted backwards by 45 degrees (or maximum if the screen is not tilted back by 45 degrees), the microphone axis coincides with the normal coming from the screen center, the microphone front end is at 50 cm from the screen plane, the microphone is directed to the screen. The load is created using the program powerMax, the brightness of the screen is set to maximum, the temperature in the room is kept at 24 degrees, but the laptop specially not blown, so in the vicinity of the air temperature may be higher. 

Performance

The laptop uses a 10th generation (Comet Lake) Intel Core i7-10870H 8-core (16-core) processor. At a TDP of 45W, its cores have a base frequency of 2.2GHz, a maximum under single-core load of 5.0GHz, and a maximum under all-core load of 4.2GHz. Practical testing under load showed that in Normal mode the processor actually consumes 45W (although we didn't see 4.2GHz even though the PL2 is 107W), and in the three "overclocked" modes it consumes up to 62W, including 58W under AI control. The integrated graphics core of the processor we, of course, did not use, for games there is a discrete gaming graphics card.

The Gigabyte Aero 15 OLED XC laptop for machine learning performed one of the best we've seen in all our testing. It's not just a very productive laptop: it's among the leaders in the overall mobile standings. Yes, the Asus ROG Strix Scar 17 is almost 11% faster, but it has the most senior mobile processor "fired up" to consume 90 watts versus 62 watts for the not so top-of-the-line Core i7-10870H in our Gigabyte laptop. Asus TUF Gaming A15 on almost older AMD Ryzen 7 4800H (we didn't come across any mobile Ryzen 9 4000-series notebooks) is also faster, but only a little bit, by 4%. However, it also consumed 45W when running, so Ryzen 4000 power efficiency leadership is undeniable. And yet, the Core i7-10870H looks very good in this comparison: it's only slightly behind the leader, and the consumption is not that much higher (especially not as much as the Core i9-10980HK). This can already be called an adequate response from Intel, though not a winning one.

The Gigabyte Aero 15 OLED XC laptop didn't make a stunning impression on us, but we liked both its looks and its features. It has a nice metal body, heavy and thick rather than the other way around, but, in general, for a model not focused on constant lugging around, the parameters are acceptable. The performance in "processor" tests is very high, the video card is also almost top-end, so it is ready to successfully solve any problems. You can increase the memory (up to 64 GB) and install a second SSD in addition to the fast regular NVMe SSD 1 TB, so the expansion options are excellent. The laptop is cooled quite effectively, but the noise at maximum modes is expected to be very high. A serious advantage for those who need it is a bright AMOLED screen with expanded color gamut. The battery life is not impressive, but it can not be called a failure. From the pleasant little things we should mention the use of a wired network controller with 2.5 Gbit/s support.

The work of Azure AI (at least in its "local" Edge AI hypostasis) didn't particularly impress us. Absolutely all modern notebooks have automatic control of coolers depending on the heat, but we have not seen a pretty picture from presentations, with rebalancing of CPU and GPU consumption depending on the current load on these components in reality. If this technology will be brought up to speed, it may turn out to be really useful, meanwhile the AI can be included just as a means for small overclocking.

Dell G7 laptop for machine learning

The laptop uses a 6-core Core i7 and a full version of GeForce RTX 2070. This - probably the best laptop for machine learning - has 16 GB of DDR4-2933 RAM and a terabyte solid-state drive. Unfortunately, at the time of writing, I did not find such a modification on sale, but exactly the same version of the RTX 2070 in Max-Q design - Dell G7 17 7700 (G717-2482) laptop for machine learning - could be taken at Moscow retail for 130,000 rubles — this is one of the lowest prices for models with such graphics and the lowest among the versions with a 17-inch screen, according to Yandex.Market.

The Killer E2500 gigabit controller is responsible for the wired network in the laptop, and the Killer Wireless—AX 1650i is responsible for the wireless connection. The updated Wi-Fi module supports IEEE 802.11b/g/n/ac/ax standards with a frequency of 2.4 and 5 GHz and a maximum bandwidth of up to 2.4 Gbit/s and Bluetooth 5.0.

Claims have arisen about the cooling efficiency of Dell G7 components. Earlier I complained that the new version frankly saved on the cooler — and we see the fruits of such savings. So, the GPU heats up to 87 degrees Celsius — this leads to the fact that the actual GPU frequency is not much higher than the stated Boost frequency. At the same time, Intel's 6-core is also heated to 99 degrees Celsius, dropping the frequency of all cores to 3.3 GHz, which is 900 MHz less than the value declared for Core i7-10750H. The fact is that the PL1 parameter for the processor in the Dell G7 17 is set at 30 watts — that is, 15 watts lower than the standard TDP indicator. Naturally, this is due to the fact that the cooling system "does not pull" two such hot chips at the same time. Consequently, Dell engineers did everything consciously.

Keyboard 

The notebook keyboard can shimmer with all the colors of the rainbow and beats into 4 zones. The keyboard can fire statically, it can pulsate, there are even stylish options such as "breathing" (that is, it slowly lights up and slowly fades out), "rainbow" and others. All the settings are in the Alienware Command Center app. As an excuse, I can only say that the interface is not too intuitive and on the fly did not figure out how to change. You have to select "Highlighting" and then check the boxes on the left side of the lines with the names of zones, and the right side will display the selected color in the palette. And I had the inertia of poking at the color on the right side. Logically, nothing happened.

The notebook keyboard gave me positive emotions. The stroke of the keys is excellent (1.4-1.5 mm), the keys are moderately elastic, separate arrow keys, which are not combined with other functions (so, PgUp, PgDn are on the sides). There is a right Ctrl, which is combined with the button responsible for the right click (the combination Shift + F10). There is a separate numeric block.

Battery life

This laptop for machine learning comes with a 240W power supply because the RTX 2070 Super graphics card wants it that way. In addition to the appearance of the laptop, the power supply unit was also updated. Visually it is thinner, but don't let this fool you. It is still the same brick. Together with the wires the weight of the adapter is without a couple of tens of grams of 1 kilogram.

The battery life of the laptop is good for a 17-inch screen. At medium brightness in economy mode in office business, it may well last 5-6 hours. Accordingly, the Dell G7 is quite possible to take to work in a cafe. I mean, you can take it from home to the cafe by car. Just to carry 3.2 kg in a backpack is not too interesting.

Performance

In terms of performance, the Dell G7 is great:

• i7-10750H processor
• RTX 2070 Super 8GB graphics
• 16GB RAM
• A super-fast 1TB SSD

A comment needs to be made. The laptop uses RTX 2070 Super 8GB graphics. There's a bit of a catch here. It's not a Max-Q design, but it's not a desktop version either, it's a mobile one. That is, the performance is lower than the desktop, but 10-15% (depending on tasks) higher than the Max-Q.

How to choose the best laptop for machine learning

As for the specifications, it is important to meet the minimum requirements mentioned above. However, if you are looking for a laptop for ML and you have the budget, it is useful to ramp up each specification to the recommended settings for ML. They are as follows:

• CPU - Intel of the 10th generation with 8 cores
• GPU - RTX 2080
• RAM - 64 GB
• SSD - 2 TB

Portability and processing power

The first thing you need to consider is portability and processing power. When talking about laptops, portability is one of the most vital factors to consider. This is because you will need to take your laptops with you for work. In such cases, ultra-thin lightweight laptops are ideal.

But besides, one of the most important things you need is processing power. You should always consider the processor you are using with your laptop. When in doubt, go for the latest-generation processor with the lowest nanoscale technology you can afford.

The speed of code compilation is important to the machine learning engineer, and this determines the requirements for the processor. It must provide high performance in Turbo Boost mode, i.e. a short-term frequency increase. The performance of a single core is also important, since many development tasks do not involve multithreading.

RAM

The next most vital thing that you should consider is RAM. Machine learning engages in heavy algorithms. In such cases, your laptop must be efficient enough to withstand such algorithms. RAM plays a huge role in helping you achieve your goal. Better RAM helps you analyze more data without interrupting the system. We recommend 32GB RAM, although you can choose something with at least 16GB.

GPU

After RAM, the most important thing to consider is the GPU. The GPU is the graphic processing unit. Machine learning spoils many neural networks. It also contains a plethora of expensive calculations. Hence, you need a powerful GPU to get the program you want. This is what you need to look for when buying the best machine learning and AI programming laptops.

Storage

Good  enough storage improves overall system performance. So, we recommend a laptop with 1TB of storage, but you can choose anything equal to or greater than 512MB.

Display and keyboard

A machine learning engineer writes hundreds of lines of small code every day, so the screen and keyboard are the most important factors to choose from. A 13-inch display won't be enough. Aspect ratio is also important: laptops with 16 : 10 or 3 : 2 screens can fit more lines.

For prolonged work with the keyboard backlight, large keys size and stroke depth of at least 1.3 mm are useful. However, the question of convenience is always subjective, so it is worth typing a couple of paragraphs of text on the laptop before buying.

In addition, it is important to avoid atypical layouts. For example, Razer laptops from before 2020 have a right Shift that is shallow and located behind the arrow pad, making it difficult to type quickly.

Dimensions and weight

When choosing a laptop, the programmer is guided by issues of convenience and portability. And although a big screen and keyboard are extremely important, you should not forget that you will probably have to carry all this around with you.

Laptops weighing more than 2 kg are uncomfortable to carry all the time. It is also worth looking at models that charge via USB Type-C. Chargers of this type are found everywhere, which will not allow you to carry a bulky adapter.

Operating System

If you need a tool for iOS programming, the MacBook is the only right choice. Also, Apple products are suitable for development on Linux servers, because macOS is based on the Unix kernel. It means that the code working on a notebook will run on a server without any problems.

In addition, there are many web development programs written for macOS which is also an important plus. Finally, optimized fonts allow the MacBook to display more lines of code than Windows laptops with the same screen height.

Battery life

Battery life is one of the most important parameters in laptops. Many people look at the battery capacity and based on it estimate how long the device will last before it runs out of power. But this is not quite the right approach.

Notebook autonomy depends not only on battery capacity, but also on the consumption of resources by internal components. Ultrabooks use energy-efficient processors and video adapters. That's why a 50-watt-hour MacBook Air can get up to 12 hours of screen time, while a 58-watt-hour MacBook Pro 13 can get up to 9 hours of screen time.

As we said before, laptops with USB Type-C charging are the best solution in terms of versatility and mobility. However, the USB Power Delivery standard has its own disadvantages, such as the inability to transfer more than 100W of power, which limits performance.

If your tasks require more computing resources, you will have to choose among models with bulky adapters and inconvenient charging connector. Also keep in mind that Windows laptops lose performance when they run on battery power, while MacBooks provide similar power from both the mains and the battery.

Outstanding new tools for Machine Learning from Dell

To obtain meaningful information from massive amounts of data, Machine Learning engineeers need extremely powerful computing resources. For the Precision line, Dell, the world's number one workstation provider, has joined forces with NVIDIA® and other leading technology providers, such as Canonical, to offer a fully integrated artificial intelligence hardware and software solution. Dell's new Data Science Workstation (DSW) product line, while offering platforms for data science, delivers the same performance and reliability you've come to expect from Dell workstations.

Maximum productivity with data for ML engineers from HP

Process the most complex datasets with speed, train data models and create visualisations with Z workstations in both desktop and notebook formats, and equip them with the most efficient software stack for data science.

Achieve new levels of productivity and data control with Lenovo

The increasingly widespread implementation of artificial intelligence, machine learning and deep learning across all industries means there is a greater need for reliable, secure and high-performance hardware solutions. Easily manage even the most complex artificial intelligence projects with the uncompromising performance, legendary reliability and scalability of Lenovo workstations. Entirely designed not to meet, but to exceed the rigorous performance requirements of demanding artificial intelligence, machine learning and deep learning workloads, Lenovo P-Series workstations are able to adapt to ever-increasing artificial intelligence needs.

Which second-hand laptop should I buy for machine learning?

90% of the work is data manipulation, the rest is selecting a toolchain and setting up the environment... When you solve real problems, creation of AI is actually a question of research and struggle with greed (the one that doesn't let you spend money in vain), because the real problems for some reason don't want to be solved by 'a single layer perceptron of 10 neurons', for some reason they want multi-megabyte networks and gigabytes of training data.

And here comes up that it is difficult (read: impossible) to search for the answer on the CPU, and the work with video cards has a lot of nuances and a limited choice of frameworks (e.g. hardware from Nvidia supports it more often).

If with a small amount of data you can use any tool, no matter how inefficient it is, even if you do something with Excel, then with a big data you have to rack your brains and look for ways to optimize data storage and processing.

I.e. knowledge in machine learning is this struggle, and not the ability to use a framework and upload images there.

p.s. choose a laptop not by its performance, but by its usability. For example, if you need it often to carry with you in a backpack - take the lightest possible, even if it will tablet pererostostok (was my model digma cite e200 with 4gb ram, awfully comfortable thing, though weak, bought for 14t.r. sorry sechs such do not do), the same heavy tasks run on the server, the network, rented or standing at your home.

If you need a laptop to stand on the table and not move, then do not buy it, better buy a desktop
 

Summary

Machine learners, deep learners, and data scientists are constantly searching for the edge on their data-driven devices. That's why we've looked at over 2,700 laptops to bring you what our team considers the best laptops for your machine learning, deep learning and data science projects. We'll continue to update this resource with new powerful and more performant machines for every budget as technology continues to evolve to deliver the best recommendations for your machine learning, artificial intelligence and data science applications.

Finding the best laptops for machine and deep learning requires research and experimentation. In our experience, we recommend starting with a laptop that costs between $900-$1500. You should consider various factors such as processor speed, memory size, GPU capacity, screen resolution, battery life, weight, keyboard quality, and overall build quality. Once you have determined the best laptop for machine and deep learning, you can then determine whether you would benefit from purchasing additional accessories such as a docking station, external monitor, mouse, and keyboard.