In this article, we discuss the types of cooling systems In the Engine. The following two systems are used for cooling the I.C engines these days:
The Necessity of Cooling System In Engine
All the I.C engines require a cooling system because the combustion of fuel takes place inside the engine itself. All the heat produced by the combustion of fuel in the engine cylinders is not converted into useful power at the crankshaft. Only about 30% of the heat is converted into mechanical work. About 40% goes off through the exhaust. The remaining 30% is useless to waste heat.
It is seen that the quantity of heat given to the cylinder walls is considerable and if this heat is not removed from the cylinders, it would result in the seizing of the piston, high fuel consumption, pre-ignition, and burning of lubricant, etc.
Keeping the above facts in view, it is observed that suitable means must be provided to dissipate that excess heat from the cylinder walls so as to maintain the temperature below certain limits. Therefore the method of removing away the excess heat from the engine cylinder is called a cooling system.
Types of Cooling System In Engine
Following are the two types of cooling systems for engines:
- Air cooling system
- Water cooling system
Air cooling system
An air-cooled system is generally used in small engines say up to 15-20 kW. The air system is used in the engines of motorcycles, scooters, airplanes, and other stationary installations. In countries with cold climates, this system is also used in car engines.
The heat is dissipated directly to the atmospheric air by conduction through the cylinder walls. In order to increase, the rate of cooling, the outer surface area of the cylinder and cylinder head is increased by providing radiating sins and flanges. In bigger units, fans are providing to circulate the air around the cylinder walls and cylinder head.
Advantages of Air Cooled System
Following are the advantages of the air-cooled system:
- The air cooled system has no radiator or pump so the system is light.
- In case of water cooling system there are leakages, but in this case, there are no leakages.
- Coolant and antifreeze solutions are not required.
- This system can be used in cold climates, where if water is used it may freeze.
Disadvantages of Air Cooled System
- Comparatively, it is less efficient.
- It is used in aeroplanes and motorcycle engines where the engines are exposed to air directly.
Water Cooling System
The water cooling system is used in the engines of cars, buses, trucks, etc. In this system, the water is circulated through water jackets around each of the combustion chambers, cylinder, valve seats, and valve stems.
The image is from https://vehiclemaintenanceandrepairs.com
The water is kept continuously in motion by a centrifugal water pump which is driven by a V-belt from the pulley on the engine crankshaft. After passing through the engine jackets in the block and cylinder heads.
The water is passing through the radiator. In the radiator, the water is cooled by air drawn through the radiator by a fan. Usually, the fan and water pump are mounted and driven on a common shaft. After passing through the radiator, the water is drained and delivered to the water pump through a cylinder inlet passage. The water again circulated through the engine jackets.
Parts of Water Cooling System
- Thermostat valve.
- Water pump
- Water Jackets.
- Antifreeze mixtures.
Types of Water Cooling System
water cooling system has the following types:
- Thermosyphon cooling system.
- Impeller circulation system.
- Pump circulation system.
- Pressurised cooling system.
1. Thermosyphon Cooling System
In this system, the natural convection of water is used. i.e hot water has a lower density and rises up. The hot water in the water jacket rises up by this way and goes radiator top, from where it passes down and is cooled, and is collected in the bottom tank.
From the bottom tank again it goes to the water jacket of the cylinder block. This natural circulation system is used in very old model cars.
2. Impeller System
This is an improvement over the thermosyphon cooling system. And it is similar in all respects except that it has a small impeller unit driven by the engine itself to increase the circulation rate of water.
3. Pump Circulation
This is the modern arrangement of water cooling in which the water is circulated effectively from the cylinder block to the radiator and back by means of a centrifugal pump driven by the engine v-belt.
4. Pressurised System
This is an improvement over the pump circulation system of cooling. In this system, there is a special radiator cap having a spring-loaded pressure valve and vacuum valve.
The cap is gas-tight when the engine is heated, water vapor is formed and this is not allowed to go out. Therefore the boiling point of water in the system is raised, so that the working temperature of the engine also rises, Higher temperature gives better thermal efficiency.
The pressure valve opens when the pressure in the system exceeds a certain predetermined valve, say 0.5 kg cm². and allow the steam to escape, thereby avoiding excessive pressure formation. When the engine is cooled the vacuum valve opens and compensates for the loss of water or air.
Advantages of Water Cooling System
- Uniform cooling of the cylinder, cylinder head and valves.
- Specific fuel consumption of engine improves by using a water cooling system.
- If we employ a water cooling system, the engine need not be provided at the front end of moving vehicle.
- The engine is less noisy as compared with air-cooled engines, as it has water for damping noise.
Disadvantages of Water Cooling System
- It depends upon the supply of water.
- The water pump which circulates water absorbs considerable power.
- If the water cooling system fails then it will result in severe damage of the engine.
- The water cooling system is costlier as it has a number of parts. Also, it requires more maintenance for its parts.
Read also: Basic Engine Components (Engine parts Names and Pictures)
Difference Between The Air Cooling System and Water Cooling System
In [Cooling system for ic engine] The following point is important for the comparison of air cooling and water cooling system.
Air Cooling System
- The design of this system is simple and less costly.
- Weight of the cooling system (per b.h.p. of the engine) is very less.
- The fuel consumption (per b.h.p. of the engine) is more.
- Its installation and maintenance are very easy and less costly.
- There is no danger of leakage or freezing of the coolant.
- It works smoothly and continuously.
Moreover, it does not depend on any coolant.
Water Cooling System
- The design of this system is complicated and more costly.
- Weight of the cooling system (per b.h.p. the engine) in much more.
- The fuel consumption (per b.h.p. the engine) is less.
- Its installation and main tenace in difficult and more costly.
- There is a danger of leakage or freezing of the coolant.
- If the system fails, it may cause serious damage to the engine within a short time.
That is it, Thanks for reading. Now, I hope that you have learned about the “Types of cooling system in engine
Silent cooling system in theory and practice
Computer user constantly struggles with noise. Anyhow, in unequal battle with a wind everyone practically loses, but there are some ways, allow to make a system block more silent. So, in this review we will try to give you some advises — what should you do with your PC to spare your nervous system.
Powerful PCs and other electronic equipment cause a constant noise disturbance in many offices. This noise pollution gets on nerves. Faster and more powerful PCs produced a significant amount of noise so, in the end of 90th years users faced with a sharp problem of noise. Loud background noise has a significant effect on the work productivity – a serious issue impacting on concentration, increased stress levels in a workplace — were the main reasons to begin using of fans. Firstly, fan placed on radiator of processor — was called «cooler». Then, small fans were also used on chipsets of motherboards. Everything was in such way. Evidently, such set of fans in case made enough noise. The peak of struggle against an irritable noise has fallen to the end of 2004 when the companies-manufacturers of cases has rushed on solving it. As a result — a huge amount of «noise problem» solutions! Some of them are relevant even now. So, we go on…
Enemy is evident — a stream of air, but for a long time there was no any efficient idea how to cope with it. I shall remind, that an idea to make a silent «box» was realized to an extreme — by the acoustic cotton wool, which was used as an internal surface of the case.
Yes, they have got a desirable effect, but after came a need of very qualitative ventilation, which depends completely on fans. Is necessary to explain, that such sound insulation maximum reduced the noise of the fan on the processor and a videocard.
But PC was still not ultra-quiet and can also reduce the quality of your entertainment and gaming experience.
You want to reduce or even to avoid noise at all? Ok, but for this your should notice the main conditions at which it will be impossible to do. We shall list them:
- close and small case;
- not enough perforation;
- dusty room (a carpet on the floor);
First point from this list above can be easy realized (read our reviews – about a great variety of correct cases in the market), but as for the rest – there can be problems.The 80mm fan makes twice loud noise than 120mm one. Also dirty fan rotates more slowly and soon can produce a very squeal sound.
Formerly, fans were lubricated, but now you will not find such type of fan as they were.
So, the right way — to change … Most opened to injury coolers are on videocard and just this place people always forget to clean up, although to blow interiors of the case by compressed air or by special vacuum cleaner is more than enough.
For sure you do not have such a special vacuum cleaner. Quite often we meet loudly roaring computers at which a cooler on a videocard can jam in any moment; in the rest everything is normal.
Cleaning of a carpet can be eternal, you can do it daily and weekly, but it will stay the same and the only what you can get will be a constant static charge on your slippers. Dust will become your constant friend! So, our advice – try to escape of carpets in the room where is your working place, where you have a computer and electronic technics.
People who tried somehow to resist a noise from system block outnumber of those who gave up at once. Mostly people bought device after device aspiring for better, however PC calmed down for a short time.
Whereby, first of all it is necessary to choose a big and spacious case for your computer. Please read a review «What case to choose» — and choose the right one.
After the model is chosen, you need to estimate the sizes of disaster, because a big system block will borrow lot of place in your apartment.
Be convinced, that in the case it is possible to place 120 mm fans and if you plan to put PC on a floor, look, is there a grid in front of the fan — the bottom part, forward panel of the case.
Two years ago we have got a system block Chieftec Dragon for testing in our laboratory.
We have equipped it with the set of Titan coolers and have started — resembled a mini airdrome on a table, as for the system block it seemed to rise up and fly away in any moment.
Coolers were absolutely new! So, make your own conclusions. Thus, we deduce a first rule for silent cooling — use only 120 mm fans.
Secondly, after fans, an optical drive and HDD can also make a noise. Though HDD are not so loud in work than modern optical drives, which are still well known for noise. Unfortunately, to cope with such misfortune you can only by means of a case with a door.
But we can sat, that such decision is not super — a noise from twisted CD only muffles a little.
As for other drives they do not rustle practically, but for this you need to pay, price is high, so for somebody such way of struggle against a rumble CD’s – does not give a best fit! Second rule for silent cooling — use only expensive and qualitative periphery. This rule is strict, because not all users can pay for a DVDRW-drive about one hundred dollars. Nevertheless, we all should aspire to it – a result will surely surpass all your expectations.
Lastly, one more fan which is located in the PSU. The most of problems are just with it. PSU can be expensive and at the same time has a cheap and noisy fan.
Replace it or not – it’s up to you! From our side, we can advise you an original solutions from FSP company(«Testing of PSU 'ZEN' from FSP»).
Probably, you know that cooling of PSU is realized exclusively by means of radiators.
Thus, the fan is closed completely — allows to avoid noise.
By the way, any processor or videocard can be also cooled not merely by fans. Think over all possible consequences in advance of the experiments. For example, top chips and processors are necessary to cool actively so, we shall not manage here without a fan.
For simple and out-of-date products – a usual radiator will be enough. Well-known Zalman company regularly launches into the market interesting coolers with a low speed or even without it at all.
Majority of motherboards come without fans, though some time back on some Abit boards has been established a small and noisy cooler.
Here we come, a third rule for silent system — use of radiators with large dimensions. In this case you relieve yourselves of fans and as a consequence sharply reduce a noise level. Attention! Such actions can damage your computer – you should realize our advices without fanaticism or otherwise you will get some problems in the future.
When everything is chosen and bought, comes a final — to collect all. Do not forget some simple laws of physics and to place fans and internal periphery due to them. So, for example, the bottom fan should absorb an air and a top one — only sniff.
These two fans should be the biggest in the whole system, because they form the basic air stream through a system block. Next, all other small fans should also absorb, because their main function is to cool HDD or radiator on a motherboard.
Finally, you can get an extremely efficiency from a purchase and installation of «reobas» for regulation of rotation speed of all coolers in a system. We can recommend you any from CoolerMaster Aerogate.
For making the process of cleaning more easier in the future, from any dust in your PC, please keep all loops inside the system block in order — accurately intercept by kapron all couplers. And the last – once a year blow the system block from a dust, brush away a dirt from fan’s blades and never forget about a videocard — here dust amass first of all.
Of course, there is no need to follow all our advices.
Creation of excellent and absolutely silent cases, even with the use of fans, continues and in this case a topic is very interesting, there are many aspects which deserve to be discussed over.
How to create such type of system blocks, we shall review next, because the necessity to have a close look to this eternal problem of cooling will be exist always!
Computer cooling — Wikipedia
Removal of waste heat from a computer
A finned air cooled heatsink with fan clipped onto a CPU, with a smaller passive heatsink without fan in the background
A 3-fan heatsink mounted on a video card to maximize cooling efficiency of the GPU and surrounding components
Commodore 128DCR computer's switch-mode power supply, with a user-installed 60 mm cooling fan. Vertical aluminium profiles are used as heatsinks.
Computer cooling is required to remove the waste heat produced by computer components, to keep components within permissible operating temperature limits. Components that are susceptible to temporary malfunction or permanent failure if overheated include integrated circuits such as central processing units (CPUs), chipsets, graphics cards, and hard disk drives.
Components are often designed to generate as little heat as possible, and computers and operating systems may be designed to reduce power consumption and consequent heating according to workload, but more heat may still be produced than can be removed without attention to cooling.
Use of heatsinks cooled by airflow reduces the temperature rise produced by a given amount of heat. Attention to patterns of airflow can prevent the development of hotspots. Computer fans are widely used along with heatsink fans to reduce temperature by actively exhausting hot air.
There are also more exotic cooling techniques, such as liquid cooling. All modern day processors are designed to cut out or reduce their voltage or clock speed if the internal temperature of the processor exceeds a specified limit.
This is generally known as Thermal Throttling, in the case of reduction of clock speeds or Thermal Shutdown in the case of a complete shutdown of the device or system.
Cooling may be designed to reduce the ambient temperature within the case of a computer, such as by exhausting hot air, or to cool a single component or small area (spot cooling). Components commonly individually cooled include the CPU, graphics processing unit (GPU) and the northbridge.
Generators of unwanted heat
Integrated circuits (e.g. CPU and GPU) are the main generators of heat in modern computers. Heat generation can be reduced by efficient design and selection of operating parameters such as voltage and frequency, but ultimately, acceptable performance can often only be achieved by managing significant heat generation.
The dust buildup on this laptop CPU heatsink after three years of use has made the laptop unusable due to frequent thermal shutdowns.
In operation, the temperature of a computer's components will rise until the heat transferred to the surroundings is equal to the heat produced by the component, that is, when thermal equilibrium is reached.
For reliable operation, the temperature must never exceed a specified maximum permissible value unique to each component.
For semiconductors, instantaneous junction temperature, rather than component case, heatsink, or ambient temperature is critical.
Cooling can be impaired by:
- Dust acting as a thermal insulator and impeding airflow, thereby reducing heatsink and fan performance.
- Poor airflow including turbulence due to friction against impeding components such as ribbon cables, or incorrect orientation of fans, can reduce the amount of air flowing through a case and even create localized whirlpools of hot air in the case. In some cases of equipment with bad thermal design, cooling air can easily flow out through «cooling» holes before passing over hot components; cooling in such cases can often be improved by blocking of selected holes.
- Poor heat transfer due to poor thermal contact between components to be cooled and cooling devices. This can be improved by the use of thermal compounds to even out surface imperfections, or even by lapping.
Because high temperatures can significantly reduce life span or cause permanent damage to components, and the heat output of components can sometimes exceed the computer's cooling capacity, manufacturers often take additional precautions to ensure that temperatures remain within safe limits.
A computer with thermal sensors integrated in the CPU, motherboard, chipset, or GPU can shut itself down when high temperatures are detected to prevent permanent damage, although this may not completely guarantee long-term safe operation.
Before an overheating component reaches this point, it may be «throttled» until temperatures fall below a safe point using dynamic frequency scaling technology.
Throttling reduces the operating frequency and voltage of an integrated circuit or disables non-essential features of the chip to reduce heat output, often at the cost of slightly or significantly reduced performance. For desktop and notebook computers, throttling is often controlled at the BIOS level.
Throttling is also commonly used to manage temperatures in smartphones and tablets, where components are packed tightly together with little to no active cooling, and with additional heat transferred from the hand of the user.
Mainframes and supercomputers
As electronic computers became larger and more complex, cooling of the active components became a critical factor for reliable operation. Early vacuum-tube computers, with relatively large cabinets, could rely on natural or forced air circulation for cooling. However, solid-state devices were packed much more densely and had lower allowable operating temperatures.
Starting in 1965, IBM and other manufacturers of mainframe computers sponsored intensive research into the physics of cooling densely packed integrated circuits.
Many air and liquid cooling systems were devised and investigated, using methods such as natural and forced convection, direct air impingement, direct liquid immersion and forced convection, pool boiling, falling films, flow boiling, and liquid jet impingement.
Mathematical analysis was used to predict temperature rises of components for each possible cooling system geometry.
IBM developed three generations of the Thermal Conduction Module (TCM) which used a water-cooled cold plate in direct thermal contact with integrated circuit packages.
Each package had a thermally conductive pin pressed onto it, and helium gas surrounded chips and heat-conducting pins. The design could remove up to 27 watts from a chip and up to 2000 watts per module, while maintaining chip package temperatures of around 50 °C (122 °F).
Systems using TCMs were the 3081 family (1980), ES/3090 (1984) and some models of the ES/9000 (1990). In the IBM 3081 processor, TCMs allowed up to 2700 watts on a single printed circuit board while maintaining chip temperature at 69 °C (156 °F).
 Thermal conduction modules using water cooling were also used in mainframe systems manufactured by other companies including Mitsubishi and Fujitsu.
The Cray-1 supercomputer designed in 1976 had a distinctive cooling system.
The machine was only 77 inches (2,000 mm) in height and 56+1⁄2 inches (1,440 mm) in diameter, and consumed up to 115 kilowatts; this is comparable to the average power consumption of a few dozen Western homes or a medium-sized car.
The integrated circuits used in the machine were the fastest available at the time, using emitter-coupled logic; however, the speed was accompanied by high power consumption compared to later CMOS devices.
Heat removal was critical. Refrigerant was circulated through piping embedded in vertical cooling bars in twelve columnar sections of the machine. Each of the 1662 printed circuit modules of the machine had a copper core and was clamped to the cooling bar.
The system was designed to maintain the cases of integrated circuits at no more than 54 °C (129 °F), with refrigerant circulating at 21 °C (70 °F). Final heat rejection was through a water-cooled condenser.
 Piping, heat exchangers, and pumps for the cooling system were arranged in an upholstered bench seat around the outside of the base of the computer. About 20 percent of the machine's weight in operation was refrigerant.
Геймеры всегда помнят о важности технических характеристик их игровых машин, но иногда забывают или не обращают внимания на архитектуру охлаждения. А ведь это одно из самых важных условий хорошей производительности игрового ПК.
Если вы хотите, чтобы ваша игровая система показывала достойную производительность, вам необходимо позаботиться о соответствующей системе охлаждения.
В противном случае вам не удастся достичь максимальной производительности во время интенсивного гейминга или выполнения других ресурсоёмких задач.
Если мы посмотрим на конструкцию любого стандартного корпуса, то обнаружим, что все компоненты расположены в одном общем отсеке. В виду такой архитектуры горячий воздух от процессора, графической карты и блока питания смешивается воедино. При этом охлаждение обеспечивается одним единственным воздушным потоком.
В отличие от стандартных корпусов, серия десктопов MSI Aegis обладает эксклюзивной системой охлаждения Silent Storm Cooling.
Она разделяет пространство системы на три отдельные зоны, благодаря чему процессор, графическая карта и блок питания охлаждаются собственными воздушными потоками.
Кроме того, в определённых моделях Aegis установлено жидкостное охлаждение, что делает эту систему не только холодной, но и тихой.
|MSI Aegis||2 зоны с раздельными воздушными потоками|
|MSI Aegis X||2 зоны с раздельными воздушными потоками + жидкостное охлаждение CPU|
|MSI Aegis Ti||3 зоны с раздельными воздушными потоками + жидкостное охлаждение CPU|
Дабы не быть голословными и подтвердить эффективность архитектуры MSI Aegis, мы решили провести ряд тестов. Результаты показали, что система Silent Storm Cooling действительно производит удивительно мало шума даже при полной нагрузке. Поэтому вы можете полностью доверять особой архитектуре охлаждения MSI Aegis, существенно снижающей нагрев системы.
Уровень шума MSI Aegis
Уровень шума MSI Aegis X
Silent way: an approach of teaching a foreign language | Skyteach
In the course of time a lot of new and updated methods of language teaching emerge. They highlight different key aspects when teaching a foreign language and each of them has their peculiarities and yet some of them are more popular nowadays than the others.
Today we’ll be talking about the “Silent way” as a way of teaching a foreign language.
Though we cannot claim that this approach is everyone’s favourite due to its uniqueness and the need of an extra effort on the part of the teacher, however, it is still a quite practiced one and has its loyal followers.
This approach was created by Caleb Gattegnoin in the 1970s. The main purpose of creating this approach was to discourage teacher talk and encourage student involvement and student talk more instead. Hence, the more the students talk the less problems with fluency occur.
The main concepts of this type of learning are:
- Learning happens and is more successful if the learner discovers and inducts rather than remembers by heart or repeats.
- Learning is facilitated by physical objects, realia.
- Critical thinking and problem solving are critical aspects to pay attention to in learning.
So, what’s this approach exactly?
As far as the delivery/presentation of the new material is concerned, this approach is very structured. Language is presented through sentences, in a sequence, from simple to more challenging. In other words, the structural pattern of the language, e.g. a sentence is presented to the learners while the learners induce the rules on their own without direct instruction.
An example would probably explain the idea of this approach better. The teacher presents the following structural pattern “Pass me the blue scarf.” The class will be asked to use this structure with each other until the teacher sees that it is assimilated.
In addition, they will be asked to change the colour of the scarf. Next one student will need to pass the scarf to another person, but they won’t happen to know the pronoun “her/his”.
Here it is the teacher’s job to jump in and offer the pronoun by letting the students practice further.
According to the topic chosen, the teacher would use rods with different colors to separate the classification of words related to grammar in context. Through the use of colors contained in charts (that a teacher must do in this method) and the separated rods, teachers can explain the classification of words.
Advantages of the method
This method allows the learners to develop independence and learn by comparing things. Students learn through interaction with one another where they not only learn independently, but also develop their team working skills.
This also helps the learners to become more responsible for their learning.The teacher’s job here is crucial in creating a type of an environment where the learners would feel comfortable to share and learn on the go.
If the environment is not favourable, it might hinder the students from communicating and interacting, which will result in a boring and a no-progress session.
The idea of modelling a new structure or item of vocabulary just once also encourages learners both to listen more carefully and then to experiment with their own production of the utterance.
Challenges of the method
Together with advantages, this method can have some drawbacks as well. Namely, students might feel scared to try out the structure without knowing the translation or the exact meaning. They might be unwilling to experiment.
Students may also feel a lack of feedback as the main task for the teacher here is to observe and come up with other tasks and exercises that would fix the interaction and the language misuse.
Finally, progress might be slow, as in contrast to other methods, the Silent way focuses on learning fewer things in a fixed period of time than covering the whole ‘material’ of the course.
In any case, this is worth a try to see how your students would respond to this method.
For more insight into the topic, please check out the videos here and here.
A Full Hardware Guide to Deep Learning — Tim Dettmers
Deep Learning is very computationally intensive, so you will need a fast CPU with many cores, right? Or is it maybe wasteful to buy a fast CPU? One of the worst things you can do when building a deep learning system is to waste money on hardware that is unnecessary. Here I will guide you step by step through the hardware you will need for a cheap high-performance system.
Over the years, I build a total of 7 different deep learning workstations and despite careful research and reasoning, I made my fair share of mistake in selecting hardware parts. In this guide, I want to share my experience that I gained over the years so that you do not make the same mistakes that I did before.
The blog post is ordered by mistake severity. This means the mistakes where people usually waste the most money come first.
This blog post assumes that you will use a GPU for deep learning. If you are building or upgrading your system for deep learning, it is not sensible to leave out the GPU. The GPU is just the heart of deep learning applications – the improvement in processing speed is just too huge to ignore.
I talked at length about GPU choice in my GPU recommendations blog post, and the choice of your GPU is probably the most critical choice for your deep learning system. There are three main mistakes that you can make when choosing a GPU: (1) bad cost/performance, (2) not enough memory, (3) poor cooling.
For good cost/performance, I generally recommend an RTX 2070 or an RTX 2080 Ti. If you use these cards you should use 16-bit models. Otherwise, GTX 1070, GTX 1080, GTX 1070 Ti, and GTX 1080 Ti from eBay are fair choices and you can use these GPUs with 32-bit (but not 16-bit).
Be careful about the memory requirements when you pick your GPU. RTX cards, which can run in 16-bits, can train models which are twice as big with the same memory compared to GTX cards. As such RTX cards have a memory advantage and picking RTX cards and learn how to use 16-bit models effectively will carry you a long way. In general, the requirements for memory are roughly the following:
- Research that is hunting state-of-the-art scores: >=11 GB
- Research that is hunting for interesting architectures: >=8 GB
- Any other research: 8 GB
- Kaggle: 4 – 8 GB
- Startups: 8 GB (but check the specific application area for model sizes)
- Companies: 8 GB for prototyping, >=11 GB for training
Another problem to watch out for, especially if you buy multiple RTX cards is cooling. If you want to stick GPUs into PCIe slots which are next to each other you should make sure that you get GPUs with a blower-style fan. Otherwise you might run into temperature issues and your GPUs will be slower (about 30%) and die faster.
Suspect line-upCan you identify the hardware part which is at fault for bad performance? One of these GPUs? Or maybe it is the fault of the CPU after all?
The main mistakes with RAM is to buy RAM with a too high clock rate. The second mistake is to buy not enough RAM to have a smooth prototyping experience.
Needed RAM Clock Rate
RAM clock rates are marketing stints where RAM companies lure you into buying “faster” RAM which actually yields little to no performance gains. This is best explained by “Does RAM speed REALLY matter?” video on RAM von Linus Tech Tips.
Furthermore, it is important to know that RAM speed is pretty much irrelevant for fast CPU RAM->GPU RAM transfers. This is so because (1) if you used pinned memory, your mini-batches will be transferred to the GPU without involvement from the CPU, and (2) if you do not use pinned memory the performance gains of fast vs slow RAMs is about 0-3% — spend your money elsewhere!
RAM size does not affect deep learning performance. However, it might hinder you from executing your GPU code comfortably (without swapping to disk). You should have enough RAM to comfortable work with your GPU.
This means you should have at least the amount of RAM that matches your biggest GPU. For example, if you have a Titan RTX with 24 GB of memory you should have at least 24 GB of RAM.
However, if you have more GPUs you do not necessarily need more RAM.
The problem with this “match largest GPU memory in RAM” strategy is that you might still fall short of RAM if you are processing large datasets. The best strategy here is to match your GPU and if you feel that you do not have enough RAM just buy some more.
A different strategy is influenced by psychology: Psychology tells us that concentration is a resource that is depleted over time. RAM is one of the few hardware pieces that allows you to conserve your concentration resource for more difficult programming problems.
Rather than spending lots of time on circumnavigating RAM bottlenecks, you can invest your concentration on more pressing matters if you have more RAM. With a lot of RAM you can avoid those bottlenecks, save time and increase productivity on more pressing problems.
Especially in Kaggle competitions, I found additional RAM very useful for feature engineering. So if you have the money and do a lot of pre-processing then additional RAM might be a good choice.
So with this strategy, you want to have more, cheap RAM now rather than later.
The main mistake that people make is that people pay too much attention to PCIe lanes of a CPU. You should not care much about PCIe lanes. Instead, just look up if your CPU and motherboard combination supports the number of GPUs that you want to run. The second most common mistake is to get a CPU which is too powerful.
CPU and PCI-Express
People go crazy about PCIe lanes! However, the thing is that it has almost no effect on deep learning performance. If you have a single GPU, PCIe lanes are only needed to transfer data from your CPU RAM to your GPU RAM quickly.
However, an ImageNet batch of 32 images (32x225x225x3) and 32-bit needs 1.1 milliseconds with 16 lanes, 2.3 milliseconds with 8 lanes, and 4.5 milliseconds with 4 lanes.
These are theoretic numbers, and in practice you often see PCIe be twice as slow — but this is still lightning fast! PCIe lanes often have a latency in the nanosecond range and thus latency can be ignored.
Putting this together we have for an ImageNet mini-batch of 32 images and a ResNet-152 the following timing:
- Forward and backward pass: 216 milliseconds (ms)
- 16 PCIe lanes CPU->GPU transfer: About 2 ms (1.1 ms theoretical)
- 8 PCIe lanes CPU->GPU transfer: About 5 ms (2.3 ms)
- 4 PCIe lanes CPU->GPU transfer: About 9 ms (4.5 ms)
Thus going from 4 to 16 PCIe lanes will give you a performance increase of roughly 3.2%. However, if you use PyTorch’s data loader with pinned memory you gain exactly 0% performance. So do not waste your money on PCIe lanes if you are using a single GPU!
When you select CPU PCIe lanes and motherboard PCIe lanes make sure that you select a combination which supports the desired number of GPUs. If you buy a motherboard that supports 2 GPUs, and you want to have 2 GPUs eventually, make sure that you buy a CPU that supports 2 GPUs, but do not necessarily look at PCIe lanes.
PCIe Lanes and Multi-GPU Parallelism
Are PCIe lanes important if you train networks on multiple GPUs with data parallelism? I have published a paper on this at ICLR2016, and I can tell you if you have 96 GPUs then PCIe lanes are really important. However, if you have 4 or fewer GPUs this does not matter much.
If you parallelize across 2-3 GPUs, I would not care at all about PCIe lanes. With 4 GPUs, I would make sure that I can get a support of 8 PCIe lanes per GPU (32 PCIe lanes in total).
Since almost nobody runs a system with more than 4 GPUs as a rule of thumb: Do not spend extra money to get more PCIe lanes per GPU — it does not matter!
Needed CPU Cores
To be able to make a wise choice for the CPU we first need to understand the CPU and how it relates to deep learning. What does the CPU do for deep learning? The CPU does little computation when you run your deep nets on a GPU. Mostly it (1) initiates GPU function calls, (2) executes CPU functions.
By far the most useful application for your CPU is data preprocessing. There are two different common data processing strategies which have different CPU needs.
The first strategy is preprocessing while you train:
- Load mini-batch
- Preprocess mini-batch
- Train on mini-batch
The second strategy is preprocessing before any training:
- Preprocess data
- Load preprocessed mini-batch
- Train on mini-batch
For the first strategy, a good CPU with many cores can boost performance significantly. For the second strategy, you do not need a very good CPU. For the first strategy, I recommend a minimum of 4 threads per GPU — that is usually two cores per GPU. I have not done hard tests for this, but you should gain about 0-5% additional performance per additional core/GPU.
For the second strategy, I recommend a minimum of 2 threads per GPU — that is usually one core per GPU. You will not see significant gains in performance when you have more cores if you are using the second strategy.
Needed CPU Clock Rate (Frequency)
When people think about fast CPUs they usually first think about the clock rate. 4GHz is better than 3.5GHz, or is it? This is generally true for comparing processors with the same architecture, e.g. “Ivy Bridge”, but it does not compare well between processors. Also, it is not always the best measure of performance.
In the case of deep learning there is very little computation to be done by the CPU: Increase a few variables here, evaluate some Boolean expression there, make some function calls on the GPU or within the program – all these depend on the CPU core clock rate.
While this reasoning seems sensible, there is the fact that the CPU has 100% usage when I run deep learning programs, so what is the issue here? I did some CPU core rate underclocking experiments to find out.
CPU underclocking on MNIST and ImageNet: Performance is measured as time taken on 200 epochs MNIST or a quarter epoch on ImageNet with different CPU core clock rates, where the maximum clock rate is taken as a baseline for each CPU. For comparison: Upgrading from a GTX 680 to a GTX Titan is about +15% performance; from GTX Titan to GTX 980 another +20% performance; GPU overclocking yields about +5% performance for any GPU
Note that these experiments are on a hardware that is dated, however, these results should still be the same for modern CPUs/GPUs.
The hard drive is not usually a bottleneck for deep learning.
However, if you do stupid things it will hurt you: If you read your data from disk when they are needed (blocking wait) then a 100 MB/s hard drive will cost you about 185 milliseconds for an ImageNet mini-batch of size 32 — ouch! However, if you asynchronously fetch the data before it is used (for example torch vision loaders), then you will have loaded the mini-batch in 185 milliseconds while the compute time for most deep neural networks on ImageNet is about 200 milliseconds. Thus you will not face any performance penalty since you load the next mini-batch while the current is still computing.
However, I recommend an SSD for comfort and productivity: Programs start and respond more quickly, and pre-processing with large files is quite a bit faster. If you buy an NVMe SSD you will have an even smoother experience when compared to a regular SSD.
Thus the ideal setup is to have a large and slow hard drive for datasets and an SSD for productivity and comfort.
Power supply unit (PSU)
Generally, you want a PSU that is sufficient to accommodate all your future GPUs. GPUs typically get more energy efficient over time; so while other components will need to be replaced, a PSU should last a long while so a good PSU is a good investment.
You can calculate the required watts by adding up the watt of your CPU and GPUs with an additional 10% of watts for other components and as a buffer for power spikes.
For example, if you have 4 GPUs with each 250 watts TDP and a CPU with 150 watts TDP, then you will need a PSU with a minimum of 4×250 + 150 + 100 = 1250 watts.
I would usually add another 10% just to be sure everything works out, which in this case would result in a total of 1375 Watts. I would round up in this case an get a 1400 watts PSU.
One important part to be aware of is that even if a PSU has the required wattage, it might not have enough PCIe 8-pin or 6-pin connectors. Make sure you have enough connectors on the PSU to support all your GPUs!
Another important thing is to buy a PSU with high power efficiency rating – especially if you run many GPUs and will run them for a longer time.
Running a 4 GPU system on full power (1000-1500 watts) to train a convolutional net for two weeks will amount to 300-500 kWh, which in Germany – with rather high power costs of 20 cents per kWh – will amount to 60-100€ ($66-111).
If this price is for a 100% efficiency, then training such a net with an 80% power supply would increase the costs by an additional 18-26€ – ouch! This is much less for a single GPU, but the point still holds – spending a bit more money on an efficient power supply makes good sense.