It’s been well established for years that if you’re an enthusiast with an appetite for processors that have more cores for your multi-threaded workloads, you’re almost exclusively looking to the Xeon platform to fulfill your requirements. In fact, Intel has been making processors with more than four cores on a single die for many a year, stretching all the way back to the first Intel Nehalem processors starting with the Core i7-970. Since 2010, the limit for consumer processors developed by Intel has been six cores with hyper-threading, a trend which the company broke for the first time in 2014 with the Core i7-5960X. The expectation since then has been that Intel no longer views restricting core count as a feature for their high-end desktop platform, and everyone anticipated a new ten-core processor derived from the Xeon platform. Well, that chip has finally arrived and it’s the fastest consumer-bound processor on the planet, but it’s also unlike anything Intel has ever launched before.
The Core i7-6950X arrives with very little fanfare on the X99 chipset and the LGA 2011-3 socket. It will work on just about any X99 motherboard on the market and only requires a BIOS update for this to work, although Intel has been seeding their board partners with BIOS images to get them ready before the CPU itself launched. The existence of this processor was not a secret – numerous benchmark leaks and specification slides made their way into the media over the course of the past year, and if you’re a die-hard enthusiast you already bought one of these in the form of the Xeon E5-2640 V4 earlier this year. Under the hood they are the exact same chip, and the only differences between them are the options that Intel chooses to keep exposed to enthusiasts who are likely to buy it.
Let’s address that elephant in the room first – pricing. Anything on Intel’s HEDT platform is frequently called “monstrously expensive”, and for good reason. It’s basically Xeon hardware on a platform sold to consumers at a reduced price, but not discounted so much that Intel or its partners make a loss on supporting it. Starting this June, Intel will be replacing the Haswell-E family with these Broadwell-E processors. Some of the replacements are actually favourable to consumers, in particular the Core i7-6800K. You get the same basic core setup and PCI-Express allocation, but base and boost clock speeds are up without changing the power limits too much.
You also get the benefit of new technologies like Turbo Boost 3.0, and support for up to 128GB of DDR4-2400 RAM. That’s an incredible amount of system memory for a consumer platform to be able to support, and it’ll go down a treat with photo and video editing professionals who want the ability to support this much RAM and have the system overclocked to within an inch of its life.
But the price of these new processors might cause angst amongst consumers who were hoping they’d be lower. The Core i7-6800K starts at $434, $40 over the Core i7-5820K’s tray pricing. The Core i7-6850K only differs by about $20 compared to the Core i7-5930K. The Core i7-6900K is a $30 surcharge at best. The Core i7-6950X, however, sits alone and unchallenged at $1,723. Pricing things per thread, the Core i7-6900K has a per-thread price of $68, while the Core i7-6950X charges a $86.15 premium per thread.
Even the Intel Xeon E5-2680 v4, a fourteen-core hyper-threaded processor, is cheaper than that, at $72.70 per thread. When anyone talks about the Core i7-6950X, they certainly can’t throw in the words, “cheap” and “value for money” into the conclusion. Intel prices it this way because they know there’s no competition at this level for the consumers. Sure, there are cheaper Xeon processors that you can order through distributors or system assemblers like Evetech, and you could buy an older dual-socket motherboard and fit two Xeon E5-2640 v4 processors on there for a whopping 40 threads and a $46.95 charge per thread, but you’re not likely to. These parts simply aren’t commonly available to consumers, and many won’t know that the option exists.
It is performance at any cost, and the cost to consumers is almost twice the price of the former Intel flagship. Converted to local currency, the Core i7-6950X will retail for about R27,000.
Turning the focus to features, Broadwell-E is interesting firstly because it’s on Intel’s 14-nanometer process, which brings significant improvements in power efficiency and clock speed. It’s fairly similar overall to the FinFET processes that have been popping up from Samsung and Global Foundries, and Intel has had so much experience building these chips on this now-mature process. Clock speeds and overclocking headroom should reach well into the 4.5GHz range with extreme cooling setups, and air cooling should be just enough to keep things under control.
Aside from some new software tricks and a chunk more of L3 cache, along with the extra cores, there seems to be nothing revolutionary about the Core i7-6950X, and that’s kind of the point. Intel’s strategy in the consumer and server markets is slow, certain improvements in quality and performance.
That software trick, as it turns out, is the latest version of Turbo Boost 3.0. The Broadwell-E family still supports the regular Turbo Boost 2.0 technology, which manages clock speed and power draw according to the workloads the cores are going through, but this time things are different. Turbo Boost 3.0 does some interesting peformance analysis of the cores and will boost a single-threaded workload much further than it would normally do with Boost 2.0, sometimes reaching well beyond the 4.0GHz range.
This is where Intel gets their single-threaded boosts from. If you’re running a single-threaded benchmark or workload like Dolphin Emulator or iTunes’ transfer and encode process, a new app designed to manage Boost 3.0 sits hidden in your application tray and monitors your CPU usage. If you’re running Dolphin, the app will take the control away from the Windows scheduler, which normally directs applications to specific cores, and shifts it over to the core with the highest per-thread performance, and overclocks it to increase performance.
You can have it do this by default, where it’ll log over time which core, or cores, is the best suited to be run at a certain clock speed, or you can set the core affinity yourself for specific workloads. This increases performance compared to the Windows scheduler which prefers to stick to official boost clock speeds, and often throws workloads around to different cores for seemingly no reason at all. You can also have the foreground application run on this core at the highest peformance level, or set up a whitelist of applications that benefit from it.
Intel may claim that there’s a 15% performance boost from running applications in this manner, but whether this actually translates into a performance boost for most applications is another entirely. Some workloads might benefit from the increased speeds early on, others might not. Anandtech’s Ian Cutress discovered in his review of the Core i7-6950X that not only does Boost 3.0 sometimes behave erratically depending on the settings made by the motherboard vendor, but it also only works on a single core for a single application at once. Some board vendors will be coming out with their own solutions similar to this application, and perhaps enable higher boost clocks for multiple cores on some lightly threaded applications.
Imagine setting up games to use up to eight threads, your browser to go up to four, and very lightly threaded applications to two. That cuts down hugely on the power consumption and heat output of the chip, and ensures that you’ll always have some kind of handle on your system’s performance. Windows 10 might get the same functionality eventually, and it’s possible that this brings similar performance boosts to older processors in Intel’s stable, as well as those from AMD.
That same benefit that enables Boost 3.0 to function also makes its way into overclocking. Just like Intel’s Skylake architecture, you can overclock one core and leave the rest at stock speeds for some benchmarks, and you can overclock some cores individually and leave others at stock, helping to improve your scores when you have the time to figure out which ones give you the beast performance at specific clock speeds. That makes finding the limits of your chip much more involved and time-consuming, and possibly more fun.
We’ll have to wait until Neo returns with another installment of the NAG OC Workshop guide to explain better than I how the new features of the Core i7-6950X work, but its safe to say that records on 3DMark, HWBot and other leaderboads are all going to be summarily smashed within the next week or so. This chip is going to raise the bar for competitive overclockers at the highest level, but it’ll also greatly reduce the amount of enthusiasts who are able to get access to hardware like this. I hope that the extreme price premium is only going to be a thing for the Broadwell-E family.