Broadwell-E arrives: Testing Intel’s 10-core, $1,700 desktop CPU

Intel's 10-core Desktop CPU

Most desktop users are going to be just fine with Intel’s standard desktop Core i5 and i7 CPUs. The company offers a wide range of quad-core chips with different levels of performance and power consumption, and there are tons of motherboards in all different shapes and sizes that offer different features for different prices. The Skylake-based Core i7-6700K is Intel’s standard desktop flagship, and for many people it offers more than enough speed.


People who need more performance (and have more cash) can look to Intel’s “enthusiast” lineup, a crop of Core i7-branded CPUs that actually have more in common with the company’s server processors than the rest of its desktop and laptop chips. Intel is refreshing that lineup today with four new CPUs based on the Broadwell-E architecture, which replaces the current Haswell-E CPUs but uses the same socket, motherboards, and chipsets.


The main thing these CPUs offer over the normal Skylake desktop parts is more cores: there are new 6- and 8-core CPUs to replace the analogous Haswell-E chips, and it’s offering an all-new, ludicrously expensive 10-core Extreme Edition CPU as well. And as is often the case, these chips showcase some technology that will eventually trickle down to the more economical processors that most people actually buy. Let’s take a look.

Meet Broadwell-E

If you’re the sort of person who pays attention to processor codenames, “Broadwell” should sound familiar. It’s a tweaked version of the Haswell architecture made on Intel’s 14nm manufacturing process (a “tick” in Intel’s now-defunct “tick-tock” schedule), and while it was originally scheduled to launch in 2014, the majority of the consumer-grade CPUs gradually launched in fits and starts throughout the first half of 2015. The server-class versions of Intel’s architectures, denoted with E, EP, and EX suffixes, typically roll out around a year after the consumer versions, and Broadwell-E follows on the heels of the Broadwell-EP Xeon launch from a couple of months ago.

Know Your Codenames
Codename and year	Process	Prominent consumer CPU branding	Tick/tock
Westmere (2010)	32nm	Core i3/i5/i7	Tick (new process)
Sandy Bridge (2011)	32nm	Second-generation Core i3/i5/i7	Tock (new architecture)
Ivy Bridge (2012)	22nm	Third-generation Core i3/i5/i7	Tick
Haswell (2013)	22nm	Fourth-generation Core i3/i5/i7	Tock
Broadwell (2014-15)	14nm	Fifth-generation Core i3/i5/i7, Core M	Tick/"Process"
Skylake (2015-16)	14nm	Sixth-generation Core i3/i5/i7, Core m3/m5/m7	Tock/"Architecture"
Kaby Lake (2016?)	14nm	TBA	Tock/"Optimization"
Cannonlake (2017?)	10nm	TBA	Tock/"Process"

Compared to the standard Broadwell processors, Broadwell-E chips include more cores, more cache, and support for quad-channel 2400MHz DDR4 RAM (faster than the dual-channel 2133MHz DDR4 supported by regular Skylake CPUs). They don’t include integrated GPUs, since servers are often run “headless” without the need for monitors or GPUs and consumers who buy these kinds of processors are usually going to pair them with at least one fast dedicated GPU. Unlike the Xeon chips, they don’t support ECC RAM, but also unlike the Xeon CPUs, they do allow for overclocking and run at higher base clock speeds.
That’s all there really is to know about the architecture itself—it’s all stuff we’ve seen many times before. So let’s focus on the rest of the package: the new 10-core option, the old chipset, and a new version of Intel’s Turbo Boost feature that’s supposed to eliminate the tradeoff between single- and multi-threaded CPU performance that you normally need to make when you go with one of these high-end chips.

Four New CPUs

The die map of Broadwell-E. Intel says the blank area on the right is used primarily in the Xeon versions of the same chip; it's switched off here.

Intel The product stack. Those are some expensive processors.

The X99 platform doesn't change, though OEMs will be coming out with new boards.

X99's connectivity is still decent, it's just missing a few things compared to Z170.

These chips get Core i7-6000-series model numbers despite being Broadwell instead of Skylake.

The i7-6950X is the king of the hill.

In Intel's eyes, Broadwell-E completes the 6000-series product stack.

Thunderbolt 3 still isn't built into the chipset, though add-in cards will let you upgrade if you want.

One of the main differences between the Xeon Broadwell chips and the Core i7 versions is overclocking.

Intel’s putting out four new chips today, three of which are direct replacements for outgoing Haswell-E CPUs. The six-core Core i7-6800K replaces the i7-5820K; both include 28 PCI Express 3.0 lanes. The 6- and 8-core 6850K and 6900K replace the 5930K and 5960X, respectively; each includes 40 PCI Express 3.0 lanes. All of these CPUs have the same 140W TDPs, and though we don’t have any of them on hand to test with, it’s probably safe to say that performance is going to stay about the same—in many cases, the architectural improvements in Broadwell will be canceled out by the slightly lower clock speeds.


The 10-core Core i7-6950X is the real star here. It has 40 PCI Express 3.0 lanes, 25MB of L3 cache, and Hyperthreading, and it does it all within the same 140W TDP as the other processors. The main downside is that its list price is an excruciating $1,723, nearly $700 (£500) more than the 8-core 6900K. (Intel never gives UK pricing, but it'll be around £1,300 or £1,400).


It’s hard not to be disappointed by Intel’s pricing. Intel’s “Extreme Edition” CPUs going all the way back to the original Pentium 4 version have cost around $1,000, and in exchange for that cash you got the best consumer-branded processor that Intel could sell you at the time. For the 10-core version of Broadwell-E, you need to spend 60 percent more money to get 25 percent more cores; you could buy a capable quad-core gaming PC complete with monitor and operating system for less than the cost of the 6950X by itself. It’s not the same bang-for-the-buck you get with the 6800K, which costs around 28 percent more than the i7-6700K (plus the price increases associated with the more expensive motherboards) but offers 50 percent more cores.


This is what it looks like when you combine a lack of strong competition with a lack of demand for high-end desktop workstations; Intel’s attentions have long since shifted to laptops and tablets, and it has been years since AMD was a credible choice for a high-end system. We’d love to see six-core-and-up processors become more common in consumer desktops, but Broadwell-E won’t be the chip that does it.

New Processors, Old Chipset

New Intel processors normally come with new Intel chipsets, but Broadwell-E CPUs will use the same X99 chipset that launched alongside Haswell-E back in August of 2014. This isn’t actually all that surprising, given that the Sandy Bridge-E and Ivy Bridge-E chips both stuck with the X79 chipset, but it means that you’re going to miss out on a handful of features that standard consumer-level chipsets like Z170 include and a handful of other high-end features that we’d like to see in such high-end workstations.


To recap, X99 provides the system with eight PCI Express 2.0 lanes, up to six USB 3.0 ports, and up to eight USB 2.0 ports (for a total of 14), up to ten SATA 3.0 ports, and an integrated gigabit Ethernet port. Different motherboard makers can choose to offer different port configurations thanks to Intel’s “Flexible IO.” For instance, some boards may choose to offer two gigabit Ethernet ports and fewer SATA ports, depending on their target audience. M.2 slots for SSDs and Wi-Fi cards are also supported.


The chipset relies on the processor for higher-bandwidth PCI Express 3.0 lanes and its memory controller. Broadwell-E officially supports 2400MHz DDR4, up from 2133MHz in Haswell-E and consumer Skylake, and processors provide either 28 or 40 PCI Express 3.0 lanes for use with dedicated GPUs, Thunderbolt add-in cards, and sometimes SSDs, depending on how your motherboard is configured. Consumer Skylake CPUs offer just 16 PCI Express 3.0 lanes, intended primarily for use with GPUs.


X99 has a few notable deficiencies compared to the high-end Z170 chipset for Skylake. Z170 includes 20 of its own PCI Express 3.0 lanes, primarily useful for GPUs and SSDs. It supports a total of 10 USB 3.0 ports and 14 USB 2.0 ports. And the faster PCI Express 3.0 lanes mean that the Direct Media Interface (DMI), which transfers data between the CPU and chipset, is also faster.


So X99 is showing its age, and if new chipsets are released for new "Kaby Lake" CPUs between now and the time Skylake-E makes it to market, the tradeoffs will become even more pronounced. You need to use this chipset to use Intel’s fastest consumer-grade CPUs, but you give up some connectivity features and flexibility. A new X-series chipset, perhaps with full 10Gbps USB 3.1 gen 2 support or even integrated Thunderbolt 3, would be a better fit for the processors.


At any rate, the continued use of X99 means that anyone using a Haswell-E CPU in an X99 motherboard today ought to be able to drop in a Broadwell-E upgrade after performing a BIOS update. Some OEMs have already issued Broadwell-E compatible BIOSes for their boards, so check your support page to make sure you have the most recent version if you’re interested in upgrading (going from a 6- or 8-core Haswell-E CPU to a 6- or 8-core Broadwell-E chip probably isn’t worth the cost, but if you decide you want more cores it might be a useful upgrade).

Performance Overview and Turbo Boost Max 3.0

Here are Intel's own performance numbers for good measure. The focus is on big round numbers that don't really take cost into account.

More cores = more performance. Skylake's newer architecture and higher clock speed mean that the 10-core Broadwell-E chip is only about twice as fast as the quad-core i7-6700K in this case.

Intel's idealized use case involves gaming while also encoding and streaming high-res video—multiple medium-sized activities, not necessarily one big activity.

All of Intel’s performance figures for Broadwell-E compare the 10-core 6950X to the 8-core 5960X and the 4-core 6700K, which makes the big round performance figures on its slides look good without mentioning the cost or the performance of the lower-end CPUs in the lineup. We don’t have the full lineup to test with, but our limited testing with the Geekbench and Cinebench CPU tests bear out Intel’s estimates: if you’re using something that can use as many CPU threads as you can throw at it (or if you’re doing several more lightly threaded things at once) this is indeed a hefty processor, but Skylake's newer architecture and higher clock speed mean that it still has an edge for more lightly threaded tasks.


The most interesting thing about Broadwell-E’s performance is a new feature that doesn’t appear in Skylake or in the Xeon versions of the same chips: Turbo Boost Max 3.0. No, it’s not some kind of energy drink; rather, it’s the latest iteration of the Turbo Boost tech you already know and love, tweaked to provide a small-but-respectable boost to single-threaded CPU performance.


Turbo Boost 2.0 is still sticking around, so let’s explain it in brief: higher-end Intel CPUs have two advertised clock speeds, a base clock and a Turbo Boost clock. The base clock is the minimum speed at which the CPU can be expected to run under a heavy, sustained workload, provided you’ve paired it with sufficient cooling. The Turbo Boost clock is the maximum speed at which the CPU is designed to perform for less demanding workloads that combine short bursts of activity with periods of low or no activity.


The amount of time that a CPU can run at its peak Turbo Boost clock speed depends on your computer’s cooling. It’s also important to note that there are other, unlisted peak clock speeds that apply depending on how many of your processor’s cores are active at once. The listed Turbo Boost 2.0 clock speed typically only applies when one core is active, but the boost clock is usually lower when two or more cores are active.
One of Turbo Boost’s goals is to cancel out the tradeoff between clock speed and cores that you’ve always needed to make with multi-core processors. But it can only take things so far—all the Broadwell-E CPUs gradually decrease in clock speed as you add more cores. This is fine for multithreaded workloads, but for single-threaded tasks that can’t keep that whole processor busy, a faster CPU with fewer cores may well beat your new 10-core, $1,700 monster.


Turbo Boost Max 3.0 takes things a step further, boosting the maximum single-core performance of these CPUs beyond the clock speed listed in the spec table. A combination of hardware built into the chip and a special driver dynamically determines the individual CPU core that’s capable of running the fastest. When the load on that processor core exceeds 90 percent (apps in the foreground or whitelisted in Intel’s utility have priority, if multiple processes are running), the process is thrown to that fast core. By default, the driver checks to see if those conditions are being met once every 10 seconds, though for our benchmarking we set that interval to one second.


Minor differences in individual CPUs mean that different CPU cores in different CPU packages will be capable of slightly different speeds. To date, Intel has just specified different clock speeds for different processors and shipped anything that could hit those promised speeds, and Turbo Boost Max 3.0 takes things one step further by squeezing out the last bit of performance from cores that are capable of exceeding those speed targets. The same technology can be used to overclock individual CPU cores, theoretically allowing you to find the best possible clock speed for each one.


At the moment, Turbo Boost Max 3.0 requires its own driver installation to work—without it, you’re restricted to Turbo Boost 2.0 speeds, which actually does make a noticeable difference for certain workloads. It’s visible in the Cinebench single-core test, which pushes the core past that 90 percent utilization threshold, but it’s effectively invisible in Geekbench 3’s single-core mode. This inconsistency will hold true for your apps, too, depending on what they’re doing.

Intel's barebones Turbo Boost 3.0 app. With some time and some luck, hopefully this will become as invisible to users as Turbo Boost 2.0 is today.

Cinebench is an example of a test that shows a clear difference between Turbo Boost 2.0 and 3.0.

Geekbench 3, not so much.

Without the Turbo Boost Max driver, the CPU’s maximum speed as measured by the Windows Task Manager was about 3.4GHz. With the Turbo Boost Max driver installed, that jumped to about 3.7GHz. Processor power consumption as measured by Intel’s Power Gadget also increased just a little depending on whether the driver was installed or not. It used about 30W of power in Turbo Boost 2.0 mode and about 34W in Turbo Boost Max 3.0 mode.
Interestingly, Intel makes no specific promises about the clock speed that Turbo Boost Max 3.0 will be able to hit. If the feature ever makes it down to quad- or dual-core CPUs, we could actually see quite a bit of variability in peak clock speeds from chip to chip. AnandTech also reports that the behavior can vary from motherboard to motherboard depending on how individual OEMs have chosen to implement it. We left all the settings on the Gigabyte GA-X99-UD3P we tested with at the default settings (the BIOS version is F22i, technically labeled as a beta).

Conclusions

The raw performance of a 10-core desktop CPU with a 140W TDP is impressive, but otherwise Broadwell-E comes with more than a few caveats.


The best-performing version of the chip costs over $1,700, much more than the other chips in the family and more than the cost of a nice gaming PC based on a more sane quad-core CPU. Turbo Boost 3.0 is a really interesting idea, and it does demonstrably improve performance sometimes, but it’s going to have some teething issues as the drivers mature and motherboard OEMs figure out how best to implement it. And the X99 chipset has fallen behind the standard desktop Z170 chipset in a few important ways that won’t matter for a lot of people but are more likely to matter to the kinds of high-end users Broadwell-E is targeting.


If you’re on the fence about buying or building an X99-based system, the 6- and 8-core Broadwell-E CPUs don’t change the value proposition much compared to Haswell-E, since the cost is about the same. Performance is better but not by a life-changing amount. Unless you want to have more cores, there’s no reason for anyone with Haswell-E to upgrade directly to Broadwell-E. And if you really want more cores but don’t care as much about single-threaded performance or overclocking, buying an actual Xeon is an option that could theoretically save you some money. 6-, 8-, and 10-core chips at low-ish frequencies will be available for $213, $306, and $667 (~£170, £250, and £530), respectively; they all support 40 PCI Express 3.0 lanes, and ECC RAM support may be a bonus depending on what you’re doing.


On the other hand, if money is no object, if you want respectable single-threaded performance, and if you want a really fast CPU that you can overclock even faster, Broadwell-E is an option for you. It’s going to find some fans regardless of the price, but it would have been nice if Intel had offered more performance at the same price as last year rather than coming up with a whole new price point for the faster CPU.