Ever since Intel debuted the "E" family of hardware with Sandy Bridge-E in 2011, these processors have formed the backbone of Intel's top-tier enthusiast offerings. While the number of cores at the top of Intel's regular consumer products has stayed steady since 2009's Nehalem -- four cores with Hyper-Threading -- the company has steadily increased the number of CPU cores it offered in the "E" family, from six with Sandy and Ivy Bridge, up to eight with Haswell-E.
The problem with Haswell-E, however, was that its core counts came at the cost of lower overall speeds compared to the Core i7-4970K. This was especially true with the eight-core version of the chip, which clocked in at 3GHz base, 3.5GHz Turbo. In these cases, a Core i7-5960X wasn't always faster than the much cheaper 4970K, despite having more cores.
The Broadwell-E die. Ten cores, 25MB of shared L3 cache.
Intel has made a variety of changes to Broadwell-E that it hopes will alleviate some of these discrepencies and create a more varied product family. Let's kick off at the high level and work our way down.
The top-end Broadwell-E is a 10-core CPU with 25MB of L3 cache (2.5MB per core). Intel holds this ratio constant in its E-class processors -- eight-core chips have 20MB of cache, while six-core chips have 15MB. This new CPU is a drop-in replacement on the X99 chipset, so if you already have one of these boards for Haswell-E, Broadwell-E should be a drop-in replacement with the appropriate BIOS flash. So let's look at the chips themselves:
The 6950X is Intel's 10-core CPU, but the price tag on this chip is going to raise some serious eyebrows. At $1723, Intel is charging 1.58x more for a 10-core CPU than an eight-core chip. The problem is, you're only adding two more cores and four more threads. Paying 60% more cash for 33% more cores isn't all that appealing, even if it's technically impressive that Intel managed to add two cores while keeping clock rates the same and without increasing CPU TDP.
The Core i7-6900K carries a $1089 price tag and the same core count while increasing clock speeds slightly (up roughly 5%). The lower-end SKUs are still six-core parts and only see 100MHz increases. This isn't exactly the enhancements that dreams are made of, though we want to note that the $412 price on the 6800K is actually quite good compared to the 6700K's $327, provided you know you'll use the additional cores.
In lieu of raw performance improvements, Intel is counting on some significant updates to its overclocking feature set to win over enthusiast hearts and minds.
Per-core overclocking is a nice touch, since it offers enthusiasts the option to test each core individually, determine which frequencies work best for which CPU cores, and then set them accordingly, with individual voltage curves. The one potentially substantial caveat to this is that how a core behaves alone and how it behaves when it's loaded alongside other cores in the system could potentially be quite different. Nonetheless, per-core frequency targets could give overclockers some intriguing tuning capabilities. Let's say, for example, that you determine four of your 10 CPU cores are capable of hitting 4GHz, six can reach 3.7GHz, eight are capable of 3.6GHz, and all 10 can run at 3.5GHz. Congratulations -- assuming you've got the cooling to handle it, your 10-core chip is now running 14-16% over stock on a voltage and frequency curve you can work with and define yourself.
It's not clear exactly what VccU stands for, but the AVX ratio offset is related to a feature Intel has offered on its Xeon processors for several years. If you've paid attention to Intel's long-term FLOPS scaling, you're aware that Intel doubled the number of FLOPS it could perform per clock when it launched AVX, then doubled that figure again with AVX2.
FLOPS per clock rating on Intel CPUs.
These continual doublings don't happen for free, however -- the 256-bit AVX2 registers draw more power, which lowers the maximum frequency Intel can support. Starting with the Xeon E5v3 family, Intel began setting lower maximum frequencies for its CPUs when they were executing sustained AVX2 workloads. It's going to give enthusiasts the option to set those offsets manually, defining different throttle points or possibly eliminating them altogether if you can handle the heat the CPU is kicking out (note that Intel's thermal trip protections will continue to function normally). Intel is also claiming that its Turbo Max 3.0 feature can deliver up to a 15% improvement compared to the Core i7-5960X's method of regulating clock speed; we'll have to confirm that in benchmark testing (our test motherboards didn't even arrive until Friday, which is why we don't have a review ready to roll for you fine folks).
Platform improvements
There are two more aspects to the launch that we want to cover. First, Broadwell-E does bump up formal support for faster DDR4, up to DDR4-2400 instead of DDR4-2133. In practice, DDR4-3200 is already available, and we suspect many enthusiasts will opt for this instead, but if you care about sticking to Intel's spec, well, things are a bit faster.
Secondly, Intel is now working with motherboard vendors to build Thunderbolt 3 support into specific motherboards that are rated for the feature, as shown below:
Thunderbolt 3 has been shipping on more systems than its predecessor thanks to the decision to swap to the USB Type-C connector. If you've got a mixed workstation environment with both Apple and PC hardware, adding Thunderbolt support to the PC side of the equation is a useful capability -- and Thunderbolt 3 is significantly faster than Thunderbolt 2, with more flexibility and raw throughput.
Early thoughts
If Intel can deliver significant clock speed improvements through Turbo Boost Max 3.0, it may boost its overall CPU performance by a larger margin than the raw clock speed figures listed here. We frankly hope this is the case, as the general argument for buying a Broadwell-E over Haswell-E isn't very strong.
There are two points to be made here: First, while there's absolutely no evidence that Intel is sitting on extra performance it doesn't want to unlock, it is fair to note that Intel has faced absolutely no competition in the high-end space since it launched the Core i7 family eight years ago. Intel's modern product stack is priced by core count more than clock speed, and it hasn't stretched itself to push core counts higher in the consumer market. From 2003 - 2006, Intel moved from one core to four. Six-core chips didn't debut for another four years, and eight-core CPUs took four years after that. If Intel had been under genuine competitive pressure, it would've rolled those improvements more quickly than it did, and for a much lower price.
But -- and this is critical -- pushing higher CPU core counts into the market doesn't mean that software will magically materialize to take advantage of those cores. According to Steam's hardware survey, 47.12% of users are still on dual-core CPUs, while 45.86% have quad-core chips. The overwhelming majority of consumer software is still quad-threaded or less.
Now, it's possible that the advent of DX12 will usher in an era where higher-end multi-core CPUs will prove themselves, AMD's Zen will offer stronger competition for Intel, and high-end desktop users will rejoice to discover their CPU investments in gaming are finally paying off. That's a pretty big jump to make from where we are to where we'd need to be, however, and it normally takes the industry 3-4 years to make a move that significant, especially when it involves API updates and fundamental engine overhauls.
If you're a workstation user who wants a 10-core CPU at a higher clock for less money than an equivalent Xeon might cost, then the Core i7-6950X is exactly what you're looking for. Most gamers are still going to better served by the Core i7-6700K, though those of you who blend workstation and gaming workloads might want to take a look at the Core i7-6800K -- while you trade off some clock speed, you still get 50% more cores for 26% more cost. It's a net gain if you use workloads that can benefit. And if you're using older hardware, like a Westmere or Sandy Bridge-E system, then the option to step up to a 10-core rig may also be extremely attractive. Ordinary gamers and users, we suspect, will do best with a 6700K or possibly something from the Kaby Lake refresh Intel is expected to launch this year.
