Choosing the CPU
Our AMD Threadripper platform system buyer's guide will help you decide on what components are best for your Threadripper system build. Threadripper and other high-end desktop (HEDT) platforms are not simple builds like many think, they require many different considerations, and a simple system configurator isn't enough to take into account different issues. There are many considerations that you might not think about before ordering your parts, such as not enough current on some high wattage multi-rail PSUs.
To make things clear, I will over cover parts of the system that could negatively impact your build. Storage, graphics, and accessories are not covered by this guide, as Threadripper will take full advantage of all of these with ease (especially with those three M.2 drives directly routed to the CPU and two x16 slots directly routed without switches to the CPU).
The first step is to choose which CPU you will be using for your build. There are currently three Threadripper CPUs, the 1950X, the 1920X, and the 1900X. We have reviewed both the 1950X and the 1920X, which are 16 core and 12 core CPUs, both have SMT, which means they will offer 32 and 24 threads respectively. The 1900X will offers 8 cores and 16 threads, but it's very close to the Ryzen 7 1800X in CPU performance, which means you are basically paying for the extra platform features (PCI-E 3.0 lanes) and CPU quality (top 5% of Zen dies).
AMD's Threadripper CPUs scale in terms of price to performance quite well, so we recommend choosing a processor that fits your budget better. While all three of the CPUs have the same Max Turbo Core Speed of 4GHz and the same maximum eXtended Frequency Rage (XFR) of 4.2GHz, their base clock speeds differ. The 1950X has a base clock speed of 3.4GHz, the 1920X has a base of 3.5GHz, and the 1900X has a base of 3.8GHz. While the 1950X and 1920X have the same L3 cache size of 32MB, the 1950X has an L2 cache of 8MB and L1 of 1.5MB and the 1920X has L2 of 6MB and L1 of 1.125MB. The 1900X has an L3 cache of 16MB, L2 of 4MB, and L1 of 768KB. We also found in our review of the 1950X and 1920X, that while the CPUs have the same TDP of 180W, the 1950X uses 10W more at full load than the 1920X, so the 1920X is a bit easier to cool than the 1950X, while the 1900X should be even easier.
The good news here is that no matter what CPU you choose, you still get ALL platform features. There is no artificial restriction to PCI-E lanes or overclocking. All the CPUs are unlocked for overclocking. Each CPU will also have the ability to run 8 DDR4 DIMMs with ECC support up to 2TB all in quad channel. CPU selection is just the first thing you need to think about, and if you want to see how we gauged the 1950X and 1920X, our review can be found here.
Choosing the Motherboard and RAM
Our Threadripper TR4/X399 Motherboard Buyer's Guide can be found here. In it, you will find out what's offered on the market. A few major things to pay attention to are VRM quality, VRM cooling, and features. Almost all the motherboards for Threadripper support the x16/x8/x16/x8 PCI-E layout and offer three x4 PCI-E 3.0 M.2 slots for the latest in M.2 storage technology. All of our motherboard reviews can be found here.
After you have chosen your motherboard, then you should choose your RAM. The reason for choosing the board before the RAM is because certain boards work better with certain memory kits. It has to do with how strong of a relationship each motherboard vendor has with each memory vendor, as they will swap hardware to ensure maximum compatibility between products and can fix bugs if they arise. Each motherboard vendor puts out a Qualified Vendor List (QVL) or certified memory list either on their webpages as a downloadable document or inside the motherboard manual.
The kits listed in these lists are fully compatible with the motherboard and the Threadripper CPUs. I would use the list for your motherboard to find a compatible kit before purchasing your memory. If you don't care for memory speed or are going for very high density (128GB+), I would stick with 2400-3000MHz for memory speed. If you are going for pure speed and are all good with 32GB or 64GB, then I would look for a kit that's 3200MHz to 3600MHz.
Threadripper does much better with single rank kits rather than dual rank kits when it comes to density and overclocking. In dual rank kits, memory chips are split into two sets, where one can refresh while the other is being used, and they do provide a bump in performance (a noticeable amount in benchmarks). Single rank kits will have a 1RxN (N is the number is memory chips, typically 8) in their specs while dual rank will have 2RxN, and most of the time single rank kits have memory only on one side of the stick while dual rank typically has chips on both.
As you can see the official AMD supported memory speeds (guaranteed) are 2667MHz with four sticks of single rank, 2133MHz with eight sticks of dual rank, 2400MHz with four sticks of dual rank, and 1866MHz with eight sticks of dual rank. These numbers might seem low, as we can easily get 32GB (4x8GB) to 3200MHz and even 3600MHz with Threadripper, but AMD can't guarantee that (and shouldn't as it's overclocking).
You should keep in mind that there are many different types of DDR4, including ECC DIMMs which are error correcting and rely on an extra chip per group of DIMMs that acts as a parity to correct any errors (flipped bits caused by things like natural phenomenon). Threadripper supports unregistered ECC DIMMs (center), dual rank DDR4 DIMMs (right), and single rank DDR4 DIMMs (left). Now, AMD has a lot of resources for tuning your RAM; you can find blog posts by AMD here and here that offer a lot of insight into single vs. dual rank, speeds, timings, AMD memory settings, and overclocking.
Samsung B-die is the preferred memory chip type for overclocking, and right now they are the easiest to overclock, and resourceful Reddit users compiled a very nice list of memory DIMMs on the market and what type of chips they use; you can find it here. I should also mention that you should check the motherboard vendor's specifications to see if they have fully implemented ECC support, as not all of them have (some will support it but not implement it). I don't care for ECC support since these days we have other mechanisms to maintain memory integrity, but I also don't work on mission-critical data where ECC is a must. I recommend a 3200MHz dual rank kit (they aren't very common) if you can find one or a 3600MHz single rank kit, I recommend Samsung B-die for both if you want to overclock.
Choose Your Cooler and Thermal Paste Method
Choosing a Cooler
There are three types of coolers for users who want to run their system 24/7; air cooling, all-in-one water cooling, or custom water cooling. There is also the fully submerged method (ambient phase change using oil) and also the phase change method where a gas is compressed, expands at the CPU, and is cooled down by a radiator (subzero phase change like an AC or chiller), but those these methods are costly and/or messy.
If you are going to overclock, I highly recommend a custom water cooling loop with at least a thick triple radiator just for the CPU. There are also triple radiator AIOs, as well as dual radiator AIOs, and AMD actually has included a mounting bracket for some AIOs.
The list of AMD TR4/X399 coolers is growing longer and longer every day, at launch, there weren't many, but now there are many. AMD supplied reviewers with the Thermaltake Floe Riing RGB 360, but I found the Corsair H115i to be just as effective if not a bit more because of the stronger stock fans.
If you already have an AIO cooler, you need to check if it's on the compatible list, and if it is, then all you will need to make it work is the bracket included in the Threadripper CPU box. However, there are also new AIOs such as the Enermax Liqtech TR4 360mm that I have pictured center right, and it works better than the Corsair H115i (center left), but it's still not as effective as a very nice (albeit expensive) custom loop.
Thermal Paste Application
I recommend buying a decent amount of thermal paste and experimenting with the best application methods. I used CoolerMaster MasterGel Maker, as it came in a decent enough size to play around with and has solid properties. If you go with IC7 Diamond or any other thick paste, it's recommended to heat it up inside its syringe in hot water for a few minutes and then apply it.
These are the different methods I used and the resulting coverage pattern.
The raw results are on the table, and I get a temperature from each die (there are two dies positioned diagonally), and then I can average them. I used Intel Burn Test with 30 loops and took a measurement three loops in and then at the end.
Here we see the results of the Thermal Paste Application Test. The results are the delta between idle and load taken with the average of the temperature of the two dies. We see that AMD's recommended 5-point method works very well, as does the two vertical line method and small dabs method. However, you shouldn't take these results as an absolute answer on what method is best, as every mount will yield different results depending on how much or how little thermal paste is applied and where. You should use the diagonal screw tightening method where you screw a bit of each screw, moving diagonally across the block, so the block mounts evenly.
Choose Your Power Supply
I measured total system power consumption at the wall with the CPU overclocked to 4GHz and at stock and compiled the results. I tested with synthetics Intel Burn Test and Furmark (max CPU and GPU load), Rise of the Tomb Raider (typical game), GTA:V (typical game), Ashes of the Singularity (very resource intensive game), Intel Burn Test alone (CPU intensive only), CINEBENCH (typical CPU rendering load), and HandBrake (typical CPU transcoding load). We can see that at full synthetic CPU load, the CPU VRM itself can easily pull over 300-350W at 4GHz (CPU pulls a bit less due to VRM), and that number will go up or down depending on the VRM quality and CPU/VRM temperatures.
I was pulling a maximum of 629W with my single GTX 1080Ti, single M.2 SSD, and 1950X overclocked to 4GHz. I would recommend an 850W PSU as a minimum by these results, and over 1000W if you are going to have more than one GPU or many drives and fans. However, wattage isn't everything, the PSU rails also matter. In the case above, the CPU at 4Ghz was pulling 300-350W from the 12v rail attached to it, and that can become an issue with multi-rail PSUs.
Multi-rail PSUs used to be very common a few years ago, and sometimes vendors didn't balance them correctly with the connections they fed. Let's say your power supply has a 100A single rail (the one on the right does), and there is a short on the motherboard that the PSU or motherboard doesn't detect. In that situation, the short would pull 100A into the system until the power supply's internal over current protection (OCP) kicks in, and that would definitely fry your system components including your CPU and perhaps your GPU. To counter this, vendors have gone and divided up that single 100A rail into multiple smaller ones, let's say five 20A rails.
If the short happened inside your motherboard and it was only allotted 20A, then perhaps only your motherboard would have died instead of your CPU and GPU and motherboard, as the power supply would have caught it at the 20A trip point. We can see that the PSU on the right has divided up a 1500W rail (125A at 12v) into six rails, two of which are 20A and four of which are 30A. The 20A rail would support up to 240W (12V at 20A) and the 30A rail at 360W (12V at 20A) before they tripped OCP on the PSU and caused a shutdown. However, you can hit those limits naturally through the CPU or GPU. The truth is you don't typically see 20A rails these days, but you should keep an eye out for them, especially on a low power (<1000W) PSU with multiple rails. If the CPU is provided a single 20A or 30A 12v rail with the 1500W PSU, it could cause issues while the 1200W PSU on the right with a single 100A rail wouldn't cause issues.
If the CPU is pulling 350W (CPU pulls from 12v only), then a 20A single rail could cause major overclocking issues, as could the 360W rail. Typically, vendors will make sure that the CPU power connectors get the larger rails, but not always, as CPUs like Threadripper haven't been common in the past few years. I recommend looking for single rail PSUs or large 30-40A multi-rail PSUs to avoid overclocking issues.
Choose Your Case and VRM Cooling Solution
Most X399 motherboards have VRMs like the ones pictured above. These motherboards don't have active cooling on their VRM heat sinks, but their VRM heat sinks are beefed up a lot and are designed to take advantage of environmental airflow inside a case. VRM overheating leads to CPU throttling, and it is one of the two most common reasons that Threadripper CPUs don't perform at their best at stock and especially when overclocked. The two VRM heat sinks pictured above do best with active airflow.
One vendor (left) went so far as to integrate a hidden fan that would help actively cool down the VRM, but as most of us know; smaller fans tend to be louder when moving more air. One vendor (right) actually includes mounting hardware so you can mount a fan directly above your CPU VRM, and that is the perfect solution for a system on a test bench.
Open air cases (such as the Thermal Take Core P5) and test benches (such as the Open BenchTable), as well as custom benches, are gaining a ton of traction because they put your hardware right in front of you where you can stare at it like a work of art. While these are glorious alternatives to a case, they can produce a huge issue; lack of active airflow over the motherboard area. Worse yet, with water cooling solutions, you don't get the same indirect airflow you might get with an air cooler.
Instead of talking about theoretical scenarios, I went ahead and actually tested out how much the system environment effects VRM heat sink temperatures. I taped a K-type probe to my Sper Scientific type-K/J thermometer which offers a very fast refresh rate and logging of maximums. For the case scenario, I tested two different air flow types; positive pressure (natural airflow pushing out of the case) and negative pressure (natural airflow pulling air into the case). I set up my Corsair 900D with seven fans, the three from the Enermax Liqtech at the top of the case exhaust hot air, and three fans at the front blow air into the case. The seventh fan is a high-pressure Enermax fan mounted at the rear of the case (you can see it in the image above), which I had blow into the case for the positive pressure test and blow out of the case for the negative pressure test.
I used the thermal probe temperatures, and the probe is touching the center of the heat sink of the motherboard pictured earlier, and doesn't move at all. I tested with no fan on a test bench, an active fan on the test bench (I used the VRM fan mount pictured earlier), negative pressure inside a closed case, and positive pressure inside a closed case. Results from VR1 and VR2 are the PWM controller's estimation of the VRM temperature (they are obviously on an offset as to not freak out the user) to make sure the trend I saw on the heat sink was correct. I found that the VR1/2 temperatures did follow the trend of temperature change that my thermal probe reported. The best way to cool down the VRM is obviously with a 120mm fan mounted above the VRM pointing right at it on the test bench, and the worst way to cool the VRM is to have no airflow on an open-air test bench/open case. However, the surprise is that there is a big difference between negative and positive case pressure inside a closed case.
With positive pressure inside the case, I found that thermal probe temperatures were actually very close to that of an active fan blowing at the VRM heat sink, and I found that with negative pressure things were much worse. All of this means you can get away more easily with a build inside a case that uses positive airflow (this is recommended by many in the industry as well). I recommend using one of your stronger fans mounted at the back/front of the case blowing in, as fans at the bottom that blow in might have their airflow obstructed by your GPU.
These results are just those of my specific setup, and they may differ for you, but it's possible to keep VRM temperatures down with solid case airflow. All said and done, you should purchase some quality fans that offer excellent airflow/air pressure/noise specifications.
PRICING: You can find products similar to this one for sale below.
United States: Find other tech and computer products like this over at Amazon.com
United Kingdom: Find other tech and computer products like this over at Amazon.co.uk
Australia: Find other tech and computer products like this over at Amazon.com.au
Canada: Find other tech and computer products like this over at Amazon.ca
Deutschland: Finde andere Technik- und Computerprodukte wie dieses auf Amazon.de