While certainly not a slouch, the i7-3930K is not the fastest CPU one can buy today, even when overclocked to 4.5 GHz. If you take a look at the article Tweakers Page – Deciding between a single or dual CPU it becomes clear that a dual Xeon solution is very costly and can not be overclocked, making it a less attractive way to go than a single i7, that delivers a much better BTFB, ‘Bang for the Buck’. So no dual Xeons.
The i7-6700 does not enter the picture because it is only a 4-core CPU with a limited 8 MB L3 cache, making it at best a sideways move but definitely not an upgrade path. Losing 2 cores and 4 MB of L3 cache is not an option. That leaves three i7 CPU’s that enter the picture, the i7-5820K, i7-5930K and i7-5960X. All three are unlocked CPU’s, so they can all be overclocked to around the same speed as the old CPU, around 4.5 GHz. Clock speed is thus not a distinguishing factor. The main differences between these 3 CPU’s are cores, 6, 6 and 8 respectively, L3 cache, 15, 15 and 20 MB respectively and number of PCIe lanes, 28, 40 and 40 respectively. That last number of PCIe lanes may not look to be terribly important at first glance, especially when using CS6 that limits one to a single video card. In the case of CC where multiple video cards are supported, this is of major importance however and can be the make-or-break point for the decision of a new CPU. Only 28 PCIe lanes, of which the first 16 are used by a single video card, leaves only 12 lanes for all the rest of the components. A single m.2 drive (more on that later) uses 4 of those remaining PCIe lanes, an Areca raid controller uses 8 more lanes, leaving no lanes for a second video card when using CC, or a Black Magic or AJA card, or PCIe based SSD cards. There simply is no room for expansion with the i7-5820K.
That leaves the choice between the i7-5930K and the i7-5960X. Since both have 40 PCIe lanes and the clock speed is not relevant due to overclocking, the distinguishing characteristics are cores and L3 cache and – of course – price. Coming from a hexa core it does not make a lot of sense to limit myself to another hexa core. The performance improvement would be too small to justify the cost of a new CPU, new memory and new motherboard. Even with the significantly higher price of the i7-5960X, that CPU is the only one to deliver a clear performance increase over the old i7-3930K with 2 extra physical cores and a 20 MB L3 cache that is almost double the old cache of 12 MB.
With the chosen CPU a new motherboard with a 2011v3 socket is neceassary, but which one? There is such an abundance of X99 motherboards available, so what model should be chosen? Initially I looked at a motherboard that has two m.2 connectors, thinking that m.2 is the way to go. Unfortunately, I only found one board to sport two m.2 ports. That was the ASRock X99 Extreme11, a very pricey board. My favorite brand and model, Asus and the WS (Workstation) model only has one single m.2 port but a much more attractive price. What is wise? To make the decision even harder, rumors about deteriorating build quality of motherboards and increasing complaints that motherboards could not in all cases sustain the enormous weight of certain CPU air coolers were getting more frequent.
In the past I have always had Asus WS motherboards and experienced great service and warranty handling by Asus. Even after more than 2.5 years, a simple RMA form is sufficient to receive a new motherboard, time to install it and have the old one picked up without any hassle or grunts. Perfect service without any cost. Admitted, this only applies to the WS series with their extended warranty. The weak spot on all motherboards, especially when overclocking with increased voltages, is nearly always the capacitors on the motherboard. Unfortunately, they usually lose their capacity only very slowly over time, and when these problems occur, it is very nice to have that extended warranty, that the more affordable models simply do not have.
Add to this that Asus – at least in my perception – is an A-brand, whereas ASRock is more a B-brand and niche player and it will not come as a complete surprise that my preference was the Asus X99-E WS/USB3.1 but what about the single m.2 connector? Is that a serious drawback? Undoubtedly a m.2 drive is extremely fast, almost 4 times faster than a SATA SSD, but currently they are limited to max. 512 GB size. You may recall that in my old ‘Monster’ I have an Areca ARC-1882ix-24/4GB raid controller, which is still going strong. It offers 28 ports, 24 internally and 4 externally, so if I were to use some of these ports for a SSD array, I could easily match or even surpass the speed of a m.2 drive and have far more storage space. The conclusion is simply that missing out on a 2-nd m.2 port is not a drawback in my case. And having USB3.1 ports on the motherboard makes it even more future proof. Having dual NIC’s which can be used in teamed configuration are a necessity with server or NAS based files and WiFi is not required.
In the worst case, if I need more m.2 devices, I can add the HP Z Turbo Quad Pro PCIe3.0 16x card, which supports 4 m.2 devices.
It is no surprise if you have read Tweakers Page – What kind of PC to use that I’m not a big fan of stock watercoolers. They are not better than good air coolers (unless extremely overclocked, think 5.0+ GHz), they don’t run quieter than good air coolers, they are more costly than good air coolers, they carry the chance of leaking on expensive parts, they are more difficult to install and maintain. The main ‘marketeer’ argument is ‘silent’, which is a lot of crap. We as Adobe users know all about marketing hype, heck Adobe invented it. Corsair and similar companies tell you watercooling is silent and not more expensive than air coolers. That simply is not true. Watercooling is replacing two rather silent fans on an air cooler with two loud fans on the radiator, plus adding the sound of the waterpump. Advocates of watercooling tell you that the main benefit is the silent operation of the video card, but they conveniently forget to tell you that you need to find the correct parts in the first place and maybe limit your video card searches to far fewer models, then you have to void the warranty on the video card to install watercooling on it, increase the size of the radiator and use more fans on that bigger radiator and you are not only spending significantly more on the extra parts, but you are still using fans in your case, you still need airflow over your m.2 drive and you still need extra airflow over the VRM part of your motherboard. There is no doubt that custom watercooling or even liquid nitrogen cooling is far better and more silent than air cooling, but at the expense of a lot more money and required expertise to install and maintain such a system. That is a different league and I don’t have the knowledge and expertise to build such a system.
That leaves air coolers, and why change a winning team. Noctua has served me well in the past and provided you clean your fans from time to time, they continue to deliver great cooling capacity over the years and the rather silent fans do not fail, at least in my experience.
There is an unwritten law about the amount of memory required in a system. I’m gonna change that. I’ll write is out explicitly:
You need at least 4GB memory per physical core.
Since I use both PR and AE it is simple to say I will go for 64 GB of fast DDR4. The choice for 8 GB sticks is obvious. The recently announced 16 GB sticks are hardly available and extremely expensive, so they are out.
I’m looking for 8 x 8GB DDR4 with a minimum speed of 2666, sold either in an octo pack or two quad packs, with timings that are the same for the first 3 figures, like 16-16-16 or 15-15-15. Within the range of memory modules found that meet these requirements, I want those with the lowest tested voltage. 1.25V is better than 1.35V. Finally, the memory modules on the short-list must be on the QVL for the selected motherboard. The final choice is a balance between price, speed and availability over here. Finally I found two quad packs with the following timings and voltage: CL16-16-16-36-2N @ 1.2V.
Coming from a GTX 680, it is clear I need to invest in a new video card. Since CS6 does not support multiple video cards, the card of choice must be a pretty fast one, at least a GTX 980, but maybe even better the GTX 980Ti, because it uses a 384 bit memory bus. The Titan X is overpriced in comparison to the 980Ti, so that one need not to be considered. Especially since one pays for the 12 GB VRAM, which is way more than I will ever need with the 4K material as my maximum resolution.
There is one particular point of interest with the GTX 900 series however, that can heavily influence the choice of video card. As you can see above, my current GTX 680 is overclocked. The GPU is overclocked by +36 MHz to 1147 MHz and my VRAM is overclocked by +349 MHz to 1851 MHz. This effectively increases the memory bandwidth from 192 GB/s to 236.9 GB/s, which is a significant overclock at stock voltage. As Bill Gehrke and others have discovered, the GTX 900 series can not be overclocked for CUDA purposes. In fact, the memory clock, which is specified @ 7.0 Gbps only runs at 5.2 Gbps in CUDA mode and can not be changed. The consequence is of course that the standard GPU clock speed and memory speed need to be the highest possible.
Looking at the specs of the GTX 980 and the GTX 980Ti makes the choice pretty easy. The memory bandwidth of the 980 is 224 GB/s, which is even lower than my overclocked 680 at 236.9 GB/s. The 980Ti has a memory bandwidth of 336.5 GB/s, a clear and sizable improvement over my old card. With the decision to opt for the GTX 980Ti made, the remaining selection process, which brand and which model is straightforward. Just select the brand and model with the highest GPU clock speed and memory speed. That means a GPU clock speed of 1253 MHz and a memory clock of 7.21 Gbps.