Intel Skylake Review

Making waves on Skylake

Skylake. The name invokes images of a serene outdoor setting, with clear water surrounded by forests, and blue skies overhead—a peaceful place where all your hopes and dreams come true. You can go swimming, fishing, water skiing, or any other number of activities. In other words, it’s summertime and we’re cooped up in front of our PCs wishing we were outside in the sun. But we digress.

Skylake is also the name of Intel’s sixth-generation Core series of processors, with the promise of higher performance and improved efficiency. For the past six months, there has been a repeated mantra among computer enthusiasts looking to upgrade: “Wait for Skylake.” The wait is now over, but does the calm surface of the water hide mysteries within its deep, or is there a tempest brewing in the sky? That brings us to the question for the day: if CPUs were boats, which one would you buy?

In an unprecedented move, Intel has launched two desktop processor updates in as many months. June gave us the Broadwell for Desktop update, aka fifth-generation Intel Core Processors, and now here we are with the sixth-generation Intel Core Processors. This wasn’t the original plan. Broadwell came late to the party, and in the process it became a mobile-focused update, launching in late 2014 primarily on hybrid devices with the Y-series parts, and then in early 2015 with the U-series Ultrabook parts. Along with the quad-core desktop launch of Broadwell in June, Intel also released “normal” higher-power notebook parts, including quad-core offerings. The original goal (several years ago) was that all of these parts would launch in 2014, but the 14nm process involved some hurdles, so things got pushed back.

Returning to Skylake, while Broadwell was delayed, Intel has pushed forward with their release plans for Skylake. In the process, things got a bit interesting. With Broadwell-DT, Intel chose to equip all of the new desktop parts with Crystalwell, the 128MB eDRAM that acts as an L4 cache of sorts. Broadwell-DT also sports Intel’s Iris Pro Graphics, with over twice as many graphics units as desktop Haswell (48 compared to 20). That means for Intel’s Processor Graphics, the i7-5775C was able to set new performance records. Meanwhile, Skylake returns to the status quo of having Intel’s second-fastest graphics configuration, which in this case means 24 EUs again (we’ll likely see 48 EU and possibly 72 EU versions later). What this means for end users is that there are certain scenarios where the i7-5775C will be faster than the i7-6700K (mostly games and graphics tests while using integrated graphics).

It’s not just processor graphics where things are interesting, though; clock speeds also give us plenty to think about. When Intel launched Haswell with the i7-4770K, it had a base clock of 3.5GHz and maximum Turbo Boost clock of 3.9GHz. Compared to that, the clock speeds of i7-5775C are a bit lower: 3.3GHz base and 3.7GHz turbo. But between Haswell and Broadwell, Intel had Devil’s Canyon, essentially a performance-tuned version of Haswell for enthusiasts. The i7-4790K bumped the base clock to 4GHz and the turbo clock to 4.4GHz. The Skylake i7-6700K ends up just short of that mark, with a 4GHz base and 4.2GHz turbo. What that means is at stock clocks, i7-4790K may still outperform i7-6700K in some tests.

The Core i5-6600K is less of a wild card, as it has a 3.5GHz base and 3.9GHz turbo—just like the i5-4690K, and slightly ahead of the i5-4670K. And as with the predecessors, the Core i5 part lacks Hyper-Threading, providing four full CPU cores but no extra “virtual” cores. It’s also priced over $100 less than then i7-6700K, so if you’re looking for a chip that can get you most of the Core i7 performance at a lower price, the Core i5 is what you’ll want.

And then there’s TDP, Thermal Design Power, or the amount of heat the chip cooling is expected to dissipate. Haswell was 84W, but Devil’s Canyon bumped that up to 88W. Broadwell-DT dropped all the way to 65W, thanks no doubt to the tame clock speeds. But Skylake reverses that trend and has a 91W TDP, the highest since the 95W Sandy Bridge launch over four years ago. (Ivy Bridge, if you’re wondering, was 77W TDP.) The last three years of mainstream processors from Intel have mostly brought about minor changes in performance, newer and potentially faster Processor Graphics, and often better efficiency, particularly in lighter workloads. The TDP ratings aren’t the whole story, but it is interesting to see the slow but steady increase over the past three generations.

Of course, the mainstream Intel platforms aren’t the only game in town. Intel also has their “enthusiast” LGA2011 platform, now in its third iteration. X79 was the initial chipset for Sandy Bridge-E, with quad-channel DDR3 support, and Ivy Bridge-E used the same platform. Haswell-E brought the X99 chipset and LGA2011-3, with quad-channel DDR4 memory support. Besides being quad-channel, the LGA2011 platforms have also offered more PCIe lanes—40 to be exact, though SNB-E used PCIe 2.0 and the latest i7-5820K is limited to 28 lanes.

Needless to say, the entire Intel desktop processor world can be confusing, and Skylake only serves to further muddy the waters. Here’s the overview of the latest generations of parts. We’ve grouped Skylake and Broadwell together, with Haswell and Haswell-E in a second table, but note that Haswell and Broadwell share a platform (LGA1150) while Skylake and Haswell-E are on different sockets.

Time to Buy a Boat

Much like buying a watercraft, determining which of the processors is “best” largely comes down to personal preference. What are you doing with your PC, how much are you willing to spend, and are you interested in power-saving features? If you’re just floating around on a lake and enjoying the scenery, there’s no need for a giant yacht or a fast speedboat. Likewise, if you’re not doing some serious number crunching or gaming, just about any modern CPU is going to be “fast enough”—that’s why laptops have been making inroads on desktops, and many businesses forgo traditional desktops entirely these days. But if you want the fastest system possible and you don’t care much about portability or size, desktops still reign supreme; however, deciding which desktop CPU is best can still be tricky.

At the high end, X99 and Haswell-E are currently the most powerful consumer platform. Sure, the clock speeds are pretty tame by default, but Haswell-E can overclock to 4.5GHz with a good CLC, and 4.2GHz is readily attainable. For pure number crunching (e.g., for video encoding), there’s no beating the extra cores you’ll get from Haswell-E, and even the six-core models will generally outperform the mainstream CPUs. The catch is that Haswell-E is more power hungry, often using 50W–100W more power than the mainstream platforms. And Haswell-E will cost more for the hardware, as the CPU, motherboard, and quad sticks of RAM all add up. Still, there’s no beating the performance.

The other platforms are far more comparable, and despite minor differences in TDP, as well as some more significant differences in features (RAM, DMI interface, etc.), performance ends up being a lot closer than you might expect. Toss in overclocking and there’s even less to write home about. Perhaps the most interesting aspect of Skylake isn’t even the CPU; it’s the chipset or PCH (Platform Controller Hub). For the past several generations of CPUs, Intel has used a DMI 2.0 interface that provides the same bandwidth as a PCI Express 2.0 x4 connection: 20Gb/s in each direction, or 16Gb/s effective bandwidth (20GT/s with 8b/10b encoding). Considering the PCH now powers multiple USB 3.0 ports, SATA 6Gbps ports, and now M.2 PCIe and SATA Express, that 16Gb/s ends up being a big bottleneck. Enter DMI 3.0.

As you might guess, DMI 3.0 upgrades the link to PCIe 3.0 x4 bandwidth, basically doubling the throughput compared to DMI 2.0 (36GT/s with 128b/130b encoding gives an effective bandwidth of 32Gb/s). And with that, Intel has dramatically increased the features of the Z170 chipset compared to Z97. Z97 supported up to eight PCIe 2.0 lanes, six SATA ports, six USB 3.0 ports, and 14 USB 2.0 ports. Z170 increases the USB 3.0 ports to 10 now, and continues to offer 14 USB 2.0 ports and six SATA ports, but it has one major advantage: it supports up to 20 PCIe 3.0 lanes. Keep in mind that all 20 lanes are still routed to the CPU over the equivalent of an x4 link, but as long as everything isn’t active at once, there won’t be much congestion. The short summary is that Z170 boards are in the works with multiple M.2 PCIe ports, not to mention we should see Z170 boards with support for four-way SLI/CrossFire using x8 connections on all four cards.

Normally, we would also have some discussion of the Skylake architecture, but Intel is holding off on revealing any details until IDF later this month, so all we know right now are the core specifications, and we’ll see in a moment how they translate into real-world performance. But if you were hoping for a major revolution in processor performance with Skylake, prepare to be disappointed. Intel’s own slides state that the new i7-6700K is only “up to” 10 percent faster than i7-4790K, 20 percent faster than i7-4770K, and 30 percent faster than i7-3770K. And that’s only if you’re running stock speeds on all the processors; again, throw in overclocking and the margins shrink even further, though the newer chips do tend to be a bit more power friendly.

One noteworthy change with Skylake is that in addition to using a new socket, Skylake has support for DDR3L or DDR4 memory. In theory, that means you could use memory from an existing system with Skylake, but that’s assuming two things: first is that you have DDR3L memory; second is that you can find a Skylake motherboard with DDR3 slots. What we’re seeing so far is that Skylake is aiming at enthusiasts, and most motherboard makers are going with DDR4 support for the time being. We might see some budget-oriented motherboards in the future that opt for DDR3 support, but considering DDR3 and DDR4 memory prices aren’t all that different (ignoring the expensive “extreme” parts), even budget boards might opt for DDR4 support. For now, with no budget Skylake processors announced (the Core i3, Pentium, and Celeron parts will come later in Q3 2015), it’s all about DDR4. Here are our test platforms:

Welcome to the Turbulent Sea of Performance

Anyone expecting Skylake to boldly go where no processor has gone before is going to be a bit perturbed by these results. The i7-4790K manages to claim victories or at least tie i7-6700K in half of our benchmarks. There are some clear victories for Skylake, but there are also some pretty major missteps. PCMark 8 appears to benefit the most, with Skylake boasting nearly a 10 percent average advantage over Haswell, but there’s a problem: Broadwell-DT is actually the overall winner of the PCMark 8 tests (which do emphasize graphics performance a bit, so that makes sense). The first pass of x264 HD, on the other hand, isn’t a great showing for Skylake, and likewise ProShow Producer 5 and the 7-zip Decompress test don’t appear to like the new architecture that much.

Of course, there’s a lot at play here. We’re testing a brand-new platform, and we haven’t had a chance to fully investigate all aspects of performance. Some of our existing benchmarks are getting old, and updated versions may show better scaling. We’re also running DDR4-3000 with CL15 timings, and Skylake might do better with lower-latency memory as opposed to higher-bandwidth DDR4. These are areas we plan to investigate in the coming weeks, but for now, the results seem to repeat the old refrain: The more things change, the more they stay the same.

Even at minimum quality, The Witcher 3 at1080p is unplayable without a discrete GPU.

Let’s move on to graphics testing. We’ll start with our integrated graphics tests, with all games running at 1080p “medium” settings. (Note that a driver bug prevented us from testing GTAV performance on Skylake, so it currently scores a “0,” but we’ve notified Intel and we’ll update the tables once we have a working driver.)

Naturally, Iris Pro 6200 with Broadwell-DT sweeps the processor graphics testing—it’s not even close. Skipping over that, we can see that Skylake is a decent upgrade to Haswell for processor graphics, but that’s only important if you’re not using a discrete GPU. In terms of performance, Skylake’s Core i7 part shows on average a 20 percent increase over Haswell. It’s enough to handle 1366x768 medium-quality gaming, but 1080p medium-quality proves to be a bit too much—only Iris Pro can really hope to cope with that resolution, and even it struggles at times.

Shadow of Mordor is another game that pretty much requires a discrete GPU.

A better plan of action is to toss in a real GPU to go along with your $350 processor, which is what we did. We used Titan X to put as much pressure on the CPU as possible, running our games at the same 1080p Ultra settings we use in our GPU reviews:

Even at stock clocks, average gaming performance with a single Titan X ends up favoring what is otherwise the lowest-clocked CPU: Broadwell-DT. But we need to put that lead into perspective; the gap between the fastest and slowest CPUs with the Titan X is less than five percent, and if we drop the i5-6600K from the list, only 2.6 percent separates Broadwell-DT from Haswell, with Skylake trailing Broadwell by 1.5 percent. Will anyone ever notice the difference when actually playing games? Absolutely not! So, if gaming is your main interest, anything from an i7-2600K to the i7-6700K should still be more than sufficient, at least for single-GPU gaming.

Looking for a Speedboat?

So far, Skylake hasn’t really shown us anything to get really excited about. We don’t know if there are some hidden gems in the architecture waiting to be unlocked, but for existing programs it’s shaping up to be yet another minor speed bump. But overclocking might give us something more exciting. Maybe.

With Haswell, we were generally able to hit 4.4–4.5GHz with good liquid cooling. Interestingly, Ivy Bridge typically had slightly higher overclocking potential (4.5–4.6GHz, maybe even 4.7GHz), and Sandy Bridge before that was often capable of hitting 4.7–4.8GHz. The combined architectural improvements might add up to a 20 percent difference between Skylake and Sandy Bridge, and stock clocks could give Skylake a 30 percent lead, but at least Haswell had to give back five percent because of lower maximum clock speeds with overclocking. So where does Skylake land?

If you heard the rumor that Skylake is more overclocking-friendly than Haswell, that appears to be true. We have the i5-6600K and i7-6700K, and using a good air cooler ( be quiet! Dark Rock 3) we were able to reach 4.6GHz with the i5 and 4.7GHz with the i7 part; a better liquid-cooling solution might gain you another 100–200MHz (and it’s also worth pointing out that Skylake boxed processors will not include a CPU cooler). If you’re using Intel’s processor graphics, we were able to boost the GPU clock by 400MHz as well, though sadly, that didn’t translate into 33 percent higher frame rates—not even close. Skylake and Z170 also bring back support for BCLK overclocking, potentially allowing for better fine-tuning of your overclock, though in practice this is mostly for the hardcore overclocking enthusiasts that want to wring every last MHz from their system.

With our current overclocking results, here’s what things look like for gaming and general performance. We’ve included our LGA2011-3 GPU test bed in the mix for good measure. (Note that our Haswell chip doesn’t generally do much more than the 4790K stock clocks, which is why it doesn’t have an OC result.)

Unfortunately, even with all CPUs running overclocked, we’re still left with a situation where no single processor emerges victorious in every test. Multi-threaded tests like video encoding and 3D rendering clearly favor the additional cores of Haswell-E, and PCMark 8 mostly agrees with that—the Work subtest being the exception, with Broadwell coming out on top. Meanwhile, gaming continues to be a close race for the most part. The i7-5930K ends up basically tied with Broadwell for overclocked gaming performance, at least with a single GPU, but the i7 Skylake is tied as well (0.3 percent behind), and even the i5-6600K is close enough that you’d never notice (1.5 percent behind). With a better-overclocking Haswell chip, i7-4790K/i7-4770K should gain another percentage point or two, resulting in what is effectively a five-way tie for gaming performance. Drat. At least the overclocked i7-6700K manages a decent lead over the other quad-core parts for most non-gaming tasks.

The good news is that overclocking is alive and well with Skylake. The i5-6600K improved performance on average by 20 percent thanks to overclocking, and the i7-6700K experiences a similar boost in performance of 17 percent. Games are still GPU limited, even with the mighty Titan X, however: the i5-6600K gained three percent over stock while i7-6700K only nets a two percent increase. Integrated graphics performance (not shown) fared a bit better, improving by seven percent on i5-6600K and six percent on i7-6700K—not enough to actually make most games playable at anything more than low to moderate quality.

Come Sail Away with Me

Those who like a calm and peaceful existence in life with few disruptions and as little change as possible are sure to welcome the presence of Skylake. It’s not a bad part by any stretch of the imagination, but performance isn’t enough to rock the boat, either. If you’re planning on building a new $1,000–$1,500 desktop PC, Skylake will be a fine choice, but if you bought such a PC in the past several years, there’s no huge reason to upgrade. The story in that sense is very much like Broadwell-DT: interesting to read about, and pleasant news for people that are happy with their current PC and are hoping to use it for a few more years.

If you’re feeling a bit concerned about this trend in computing performance and need to vent, go take it out on the laptops, tablets, and smartphones of the world. Many people have discovered that tablets and smartphones are fast enough, and what’s more they’re highly portable. The number of people who don’t have a decent smartphone in the United States, especially in the 40-and-under age group, continues to shrink. Couple that with the Internet as a source of entertainment, and we have the makings of a great, big plateau rising up out of our Skylake.

Then there’s that whole Moore’s Law business, which is becoming increasingly less of a law and more like a dream of a bygone era. As with all exponential growth situations, it eventually had to slow down, and we’re seeing that in a very big way. It’s crazy to think about the 386 many of us knew in our youth; in thirty years, processor performance went up about 5000X (give or take). But in the past five years, focusing on the mainstream Intel platforms, performance hasn’t even doubled. Ouch. Adding more cores isn’t a panacea either. Sure, six-core and eight-core chips can outperform quad-core offerings in quite a few benchmarks, but there are also many cases where performance stops scaling beyond four cores. Double ouch. While processors continue to get smaller, cooling the processors is becoming more difficult, so that newer chips in often add new features rather than boosting overall performance.

From our vantage point, it looks as though Skylake and the Z170 platform end up being far more about features than performance or efficiency. Many of the Z170 motherboards will feature USB3.1 support with type-C connectors, USB 2.0 ports are hardly necessary now that we can have 10 USB 3.0 ports, and with 20 PCIe 3.0 lanes available, we can start thinking about things like multiple M.2 NVMe SSDs. These are all good things, and if you’re in the market for a new desktop, the only real question is whether you should build a Skylake-based system, or should you go with Haswell-E?

Much like choosing what sort of vessel you want to take out on the water, there’s no “best” solution that applies to everyone. If you want a speedboat with enough power to pull multiple skiers, Haswell-E remains the better choice: 40 PCIe lanes mean you can run two GPUs (or three or four GPUs) and still have spare lanes for networking, audio, M.2, and USB ports. It might make too much noise to be a good fishing vessel, but that’s okay. Meanwhile, Skylake is for smaller groups that don’t need to haul ten people with several more riding the wake. It can handle gaming nearly as well as Haswell-E, and it can go places where a bigger vessel would run aground. It’s also newer and has features you won’t find on the brute force X99 platform… at least not until Skylake-E shows up in about a year.

Follow Jarred on Twitter.