Last week I built a new PC! My old home workstation was using a Core i7-2700k and it just wasn’t cutting it for some of the VR titles. On the one hand the old CPU was 4.5 years old and that should have been plenty of lifespan. By the old Moore’s law the new CPUs should be 8x faster! Unfortunately we aren’t still on that same curve. From the straight specs the CPU only progressed from 3.5ghz to 4.0ghz (with a max turbo of 4.2ghz) but overall depending on which benchmark you read you get an overall 27-45% boost in performance and probably a bit more headroom if you count overclocking.
Before I got into the details of my new workstation I wanted to mention that whenever I give advice on a PC build its always important to ask the person what they plan on using the PC for. If you don’t have any real needs for graphics you can actually do great with some of the tiny PCs- for example the Intel NUC PCs are the size of a Mac Mini, cost ~$400 and can contain a Skylake (up to the latest generation core-i7) CPU.
Even if graphics/gaming is a big part of the use there is a pretty big spectrum of capabilities. For 90% of people building a nice small Mini-ITX machine with the new NVidia GTX 1070 is probably perfect. On the other hand if you want to drive games on a 4k monitor or do really high quality VR it might even make sense to prep for a system with SLI (dual graphics cards). You can have plenty of gorgeous VR experiences on a single GPU like last generations GTX 970 and indeed that is the target most games are aiming at now. But missing frames SUCKS. And even more VR can really be sensitive to aliasing and related issues. There are a lot of factors, but a good chunk of the root is just that can you move your head with such precision (and you can’t keep it still) so many problems that traditional video games avoid because they control the camera show up. So this is just a long way to say that while I love tiny SFF machines I designed my new workstation to support SLI and hope to at some point equip it with dual-GTX 1080s.
Having committed to supporting SLI at some point that pushed me to a mATX motherboard instead of an ITX one. In some ways its kind of a shame- I’m not aware of any technical reason why someone couldn’t build a daughter-card for an ITX motherboard that had dual GPU slots and let you position them in some sensible way in the case. But I’ve searched and that thing doesn’t appear to exist.
The other issue that SLI raises is CPU family & chipset (this gets a bit complicated). The basic dilemma is that the Skylake processors like the 6700K use the Z170 chipset and the LGA 1151 socket which provides 26 I/O lanes. These are the high speed communication between the CPU and other devices. Modern GPUs all go in PCIe slots that have 16 lanes. Of course you can do the math pretty easily and tell that if there are two GPUs and only 26 I/O lanes they aren’t each going to be able to use all 16. On the other hand the X99 which supports the LGA2011-v3 socket supports 40 lanes and so you can have dual GPUs each with 16 lanes on those systems. Unfortunately the CPUs for those sockets lag the mainstream ones by a bit- the new CPUs for the LGA2011-v3 socket called “Broadwell-E” should be releasing in a few weeks. There will be options that support 6, 8 and even 10 cores although those get incredibly expensive. Unfortunately as PC consumers we are left with trade-offs. Broadwell-E is a generation older and has lower frequencies (the 6-core one will run at 3.6ghz) which means less single-threaded performance. Of course they can make up for it by having a bunch more cores, but many graphics applications are ironically poorly threaded on the CPU-side. And they have those extra lanes so when they are running dual-GPUs they each get the full set of I/O.
But do the full 16 channels really make a difference? It took a bit of searching but I found some benchmarks from Puget Sound Systems and from TechPowerUp. To skip to the conclusions the difference between running a GTX 980 at 8 lanes vs 16 is tiny. Dropping down to only 4 lanes or all the way down to PCIe 1.1 is measurable but the difference between x8 and x16 seems within the error-bars (PCIe 2.0 vs 3.0 also didn’t appear to make much difference). The one caveat I’ll mention is that this is looking at FPS during the execution of a game. Ideally a well-written game shouldn’t be putting much load on the PCIe bus during runtime because hopefully all the big textures and such are already on the GPU. I do wonder if this would have a measurable impact on level-load time and/or games that are seamless when they have to load new content. With a PCIe 3.0 8x GPU having 4GB/s bandwidth it could take up to two seconds to swap out the entire 8GB of VRAM on these new cards. But another way to put it is that is 44MB per frame at 90fps so hopefully a game can stage loading those assets without glitching.
The other reason I wanted to go with the Skylake processor / Z170 chipset was I really wanted to get a PCIe M.2 drive for my boot drive and a USB-C port. Some of the X99 motherboards support these but as they are slightly older I was concerned the support wouldn’t be as good. As it was I had to upgrade the BIOS on the motherboard I got for the M.2 to work correctly.
One other note on the build- I really wish I could have built this using the Cerberus case from Kimera Industries. It looks like an amazing design of a small case (smaller than most ITX cases) that can still fit a mATX motherboard. Unfortunately their kickstarter didn’t get funded enough and they are still trying to figure out how to get production going. I ended up going with the Corsair Carbide Air 240 case which I have to say I’m very happy with so far. Good construction and all the panels came off which made it much easier to install components than any case I’ve ever worked with before. I also like that it keeps the power supply and drives in a separated section from the motherboard- I’m sure that helps with cooling a bunch but its even better for cable management. Of course its pretty big compared to the Cerberus (70% bigger!) but its not actually much bigger than the ITX case I had before.
So with all those details out of the way here is my current build-
Case- Corsair Carbide Air 240
CPU- Intel Core i7-6700K (4.0ghz LGA 1151 91W)
CPU cooler- Corsair Hydro Series H50 1200mm Quiet edition (get the CPU cooled by a nice quiet 120mm external fan rather than the usual coolers that sit inside the case heat)
Motherboard- ASUS ROG Maximus VIII GENE Z170 mATX
RAM- G.SKILL Ripjaws V 32GB (2x16GB – I have 4 slots so I can upgrade to 64GB later if I want). DDR4 3200 CAS14. Its not clear if fast RAM is really worth it at all but if you do look for fast RAM don’t ignore the CAS spec.
Boot drive- Samsung 950 Pro M.2 512GB (really wish the 1tb version was out but it isn’t yet).
Development drive- Crucial 1TB M500 (transferred from my old PC)
Power Supply- EVGA 850W 80+ Gold fully modular. I like fully modular for cable management and this has enough juice for SLI. Otherwise I probably would have gone with 600-650W.
TBD- NVidia GTX 1080 (possibly x2)
I also threw a pair of spinning-rust drives in there. I had been using a few external 8TB “backup” drives as normal drives for a bit but decided for the storage I’m using for all my photos, videos, and random other stuff its probably better to use drives with a somewhat higher spec. I went with Seagate Enterprise NAS HDD 8TB 7200RPM drives that I got on Amazon. The 8TB “backup” drive is quite a bit cheaper and to be honest has worked great so far. But having had an opportunity to hear a talk from a guy working for one of the drive companies I suspect some of the basic components upgrades to invisible things like the rotors, bearings, actuators and such should help with reliability.
I’ll try to post again soon with my more normal build” recommendations. Also in a couple of weeks when Broadwell-E really comes out we can start dreaming about that 8-core system…