How to pick components for a high performance desktop PC
In this, my very first blog post, I’ll walk You through how to pick components, or rather exactly what components I recommend, when building a high performance desktop PC for (GPU-based) 3D-rendering, machine learning, virtual reality, video installations/projection mapping, video editing, (certain) software development, and similar (heck, even password cracking). Ok, here we go!
GPU and GPU-cooling – Graphics Processing Unit
Probably the most common question; What GPU should I get? Short answer, pick the best one from Nvidia that you can afford right now (at least if your software of choice is built on Nvidia’s proprietary CUDA platform, otherwise you can opt for AMD). Here’s where you should really spend! The 2080 Ti is the highest performing one, while the 2070 Super has the best price to performance ratio, and runs cooler, on less power. Generally, a higher number of CUDA cores means faster computing/shorter render times, since more calculations can be done, in parallel.
(There’s the Titan RTX – “The fastest PC graphics card ever built” with 24 GB VRAM, but it’s costs near 3,000 EUR – double the price of a 2080 Ti, but only has 256 more CUDA cores; 4,608 vs. 4,352. So, Titan RTX is only useful for those who really need those 24 GBs of VRAM, e.g. larger scenes with many/detailed 3D objects.)
You might still get more ‘FPM’ (Frames Per Money) if you spend more on a higher end GPU or two (again, 2080 Ti), since the rest of your system is quite expensive and normally can host at least two (2) GPUs. Also a lot of software do scale more or less linearly when adding one or two GPUs (ex. Octane). Meaning you could get 100% more performance by spending 50% more money, effectively lowering the price per rendered frame. For VR though, a single-GPU setup is the way go, since most if not all VR-software is built to utilize only one GPU (makes no use of SLI/NVLink/Crossfire).
Note that for most of us, Nvidia’s Quadro-series is not really a reasonable option. They are tuned slower than gaming-GPUs, and have drivers mainly made for stability and longevity, while gaming GPUs are built and tuned for absolute top speed, and trust me, you do want to go for speed! The one thing that Quadro has is large amounts of (very expensive) VRAM, similar to your system RAM but for video. Only go this route if you really need more than 11 GB of VRAM (for relatively large 3D-scenes), and cannot rely on out-of-core-geometry (where the rendering software starts using CPU-RAM after GPU-RAM/VRAM is filled up). Or look up NVLink, which enables sharing mutli GPU-VRAM as one large pool (1+1=2).
Good cooling is absolutely essential to avoid throttling, i.e. GPUs slowing down to avoid destroying themselves from overheating! Different types of GPU-cooling;
– Air; Suitable for up to two (2) GPUs per build, if there’s a double-slot space in between, and it’s a well ventilated case. This is what most gaming GPUs have, recognizable by their big fans and heatsinks. Runs quiet, but heats the inside of the case. You can’t go wrong with Asus, Gigabyte, EVGA or MSI.
– Blower; Suitable for multiple GPUs sitting next to each other in one build. If you go for three or more 2080 or 2080 Ti, it’s worth considering hybrid water AIO, or even custom loop if you know what you’re doing. Blowers are recognizable by their square look and only one relatively small fan. They blow the hot air out of the case and are a bit smaller, but runs about 10% slower and is noticeably louder. Asus and Gigabyte makes good blowers.
– Water /hybrid /AIO; Suitable for multiple GPUs sitting next to each other in one build. AIO is short for All In One; a closed one-size-fits-all water loop by tubing, pump, radiator and fan. An upside with AIOs is that you get actual water cooling, that runs the GPU at lower temps, is quieter than air and blowers, and to that at higher clock speeds. Downsides? They cost about 200 EUR more per GPU, takes up more space, and is somewhat risky (water leaks). MSI and EVGA has hybrid models.
– Air-to-water conversion; There are D.I.Y. kits to transform air cooled GPUs into water /hybrid /AIOs, using universally designed parts and stock CPU-coolers. Kind of fiddly though, might involve a Dremel. Search and watch at YouTube to get an idea. Look at NZXT Kraken G12 for example.
– Water /custom loop; D.I.Y. custom water loops with blocks, tubing, pump, radiator(s) and fan(s). This is an expensive (!) investment but with a possibly high ROI (Return Of Investment). It’s more risky (water leaks), and therefor not for the faint of heart (consider avoiding this option for critical operations). But you can stack up to seven (7), or even eleven (11) GPUs in one case (motherboard dependent, server grade), still overclock them quite a bit, and by that get really-REALLY-high performance, i. e. an excellent ROI, given that you have real use for it.
This is a small physical connector for GPU-to-GPU, similar to SLI but newer and much higher performing. It enables two GPUs VRAM to be used as one shared pool, i.e. two (2) 2080 Tis, with 11 GB VRAM each, will enable a total of 22 GB VRAM for your application to use. Also, your GPUs will perform a little better speed wise, especially for machine learning. (Some software have been reported to perform worse with NVLink, so remember to do some research regarding Your software of choice).
Motherboard – The PCB board where everything else is mounted
Any popular ‘gaming’ or ‘workstation’ motherboard will usually get the job done, with up to two (2) GPUs. Though, the options are few and far between if you want to build with three (3) or more GPUs. My recommendation is the Asus WS Z390 Pro, since it can host up to four (4) GPUs running at x8 PCIe speed, with a 16 Lanes only CPU, thanks to the PLX chip. In the context of GPU rendering, PCIe x8 has no perceivable loss compared to x16, approximately only 1-2% in most cases (though, when gaming it can have a bigger impact).
Another reason to pick this motherboard is the fact that has the 1151 socket, meaning you can choose from a wide range of reasonably priced and high performing consumer/gaming-CPUs, incl. the i9-9900K. This motherboard also hosts two (2) NVMe M.2 SSDs at theoretical speeds up to 4,000 MB/s! (You can also look up the Asus X299 series as an alternative, beware they come with a different socket though.)
CPU and CPU-cooling – Central Processing Unit
You don’t really need a high number of CPU-cores, six to eight will do fine. Most software are single threaded anyways. The one situation where you do want a lot of CPU-cores though, is if you do CPU-based 3D-rendering (perhaps Arnold or V-RAY, but they are or will soon become CPU/GPU-hybrid). So, what most readers will need is rather a high clocked CPU! You absolutely want to avoid going cheap on the CPU, because an under powered one will bottleneck the GPU(s) you just invested in. For example Redshift do use the CPU a lot (ex. when loading the 3D-scenes into VRAM). The gaming flagship model from Intel, i9-9900K, can turbo speed (theoretically) up to 5.0 GHz on a single core (in reality around 4.7), which again, really helps speed up software like Redshift, Octane, Blender (Cycles, Eevee), Unity, Unreal, etc. A CPU with only 16 Lanes is not a problem in a multi GPU build, if coupled with a PLX equipped motherboard, like Asus WS Z390 Pro.
Cooling is essential, since too much heat will make your CPU throttle, i. e. slow down, if/when it reaches a certain temperature. Go for a large Noctua air cooler, or a two-fan Corsair AIO water loop. In some cases the non-water Noctuas outperform AIOs from Corsair, NZXT, and others, but bear in mind that Noctuas are quite heavy and therefor can break/detach from your CPU if transported (actually Noctua air cooling has fluid inside, but in a different way).
SSD – Solid State Drive
Rotating hard drives deliver, on average, roughly 128 MB/s read speed, and 120 MB/S when writing. (Though there are faster ones). This is why most people opt for SSDs nowadays, which can reach around 550 MB/s read speed and writes around 520 MB/s. Clearly SSDs are favorable (though still significantly more expensive), at least for hosting your Operating System, and if you’re working with large files, such as video. But for storing large amounts of data, most people and businesses still use “old school” rotating hard drives (don’t mix up desktop-drives and NAS-drives).
For a high performance build though, I recommend a new type of SSD known as NVMe PCIe SSD, which reads around 3,500 MB/s, and writes around 3,300 MB/s. Yes, that’s six, almost seven times faster than a SATA-SSD! Not only useful for hosting OS and moving large files, but also when caching a 3D simulation, or when handling high numbers of small files in general, such as when deploying software builds (ex. to a HoloLens, or an Oculus Quest), or working with software development in general. I recommend the Samsung 970 EVO Plus (1 TB).
PSU – Power Supply Unit
This is a component a lot of people don’t realize how important it is. Especially since you’re going to run high performance applications, and probably in professional context. You want enough power, and you want that power stable. For a system with a single 2080 Ti, it’s recommended to have at least a 700 Watts PSU. If you add a second 2080 Ti (TDP is 250 to 300 Watts), then you need at least a 1200 Watts PSU, since they run best at around 50% load. Modular means you can detach cables you don’t need, which ease the air flow, makes it easier to build with, and simply looks better. Corsair, Thermaltake, and EVGA are good brands, and Super Flower, which has one at 2,000 Watts!
Case aka. Chassis – The box encapsulating all components
When choosing a case, pick one that has the same form factor as your mother board. Also, make sure to go for one that is known to be able to keep relatively cool under load. Corsair, Fractal Design, and NZXT are examples of good brands, but your options are many.
RAM – Random Access Memory
This is the computer’s working memory. Maxing out RAM is probably the easiest and cheapest way to speed up a computer in general (given that you already have the OS on an SSD). Think of it this way, if a faster CPU can juggle faster, then more RAM helps your computer continue juggle at that same speed, but with more items at the same time. Like a clown who is juggling with first three, then five balls.
Most RAM today runs at a default speed of 2,666 MHz, when not overclocked. Overclocking RAM can introduce system instabilities, and the performance results vary. People report from, none at all, to 100% speed ups in render software. Personally I still run mine at stock 2,666 MHz. CL, short for CAS-latency, should be 16, 15 or even 14, where CL14 means this RAM will be slightly faster, but generally not perceivably so.
If you’re on a tight budget, go for minimum 16 GB. Otherwise, go for 32 or 64 GB. Make sure to fill all slots on the motherboard, and buy all sticks at the same time, from the same batch. Make sure they are identical! Corsair, Crucial and G.Skill are good brands.
OS – Operating System
Windows 10 Home does the job for most people. Though, if you want to be able to do some fancy stuff like remote access to your PC (use it from distance, from another address), or use the HoloLens emulator, then go for Windows 10 Pro (Edu and Enterprise versions also has all the Pro features). ‘OEM’ means it will be stuck on that one motherboard forever, while ‘Retail’ means you can easily move that license to another computer.
If you’re a macOS/Mac OS X-user, I salute you for having read this far – welcome to the dark side! If you’re a Linux-user, most of the above information should apply to you as well, but keep an eye on driver-availability for the specific components you’re planning to use. If you’re a Hackintosh-user, I bow for you, your perseverance is outstanding! For a Hackintosh-build it is EXTREMELY important that you make sure beforehand, that all parts have drivers available. But you already knew that.
Bonus! UPS – Uninterruptible Power Supply
Get a UPS for your desktop, your laptop already has one built in. Basically it’s a glorified battery, a box to sit between the wall-outlet and the PSU. If there’s a power outage, your computer will keep working as if nothing happened, instead of crashing (avoid breaking a long render). Hopefully the power will come back on before the battery is empty. You can also connect the computer and UPS via Ethernet or USB, and make use of fancy stuff like automatic and smooth powering off the computer. APC is the top brand here.
Enjoy your computer, and see You in cyber space!
How to pick components for a high performance desktop PC