We’re reviewing our first Core i9 chip—Intel’s Core i9-7900X 10-core part—as a veritable CPU storm looms. Sure, Core i9 blew in as the most powerful CPU the company has ever sold to consumers, and it’s currently the fastest Core-series CPU available. But an ill wind is blowing: AMD’s Ryzen 5 and Ryzen 7 chips offer stiff competition in the low end, while its massive 12-core and 16-core Threadripper CPUs loom on the high end.
There’s so much to say about Core i9, we put the prices, features and FAQs into a separate story you’ll want to read for background. For this review, we’ll walk through some of the under-the-hood issues directly related to performance, and then we’ll dive into the benchmarks.
Core i9 is the first new “Core i” Intel has introduced in 10 years. The company guarded the secret so closely that it even intentionally mislabeled the chips (including our review sample) as "Core i7" to throw off leakers. In fact, the review sample you see here still identifies itself as Core i7 rather than a Core i9.
Like most major Intel launches, the Core i9 family represents a new platform, not just a new CPU, which means a new chipset, the X299, and a new socket, the LGA2066, all incompatible with previous CPUs.
The new platform also does something no previous one did by unifying two CPU families. Before today, if you wanted the company’s latest Kaby Lake core, you had to buy a motherboard using the LGA1151 socket. And if you wanted to buy, say, a 6-core Skylake CPU such as Intel’s Core i7-6800K, you had to buy an LGA2011 V3-based motherboard.
With X299 and LGA2066, you can now pick your poison, because the platform encompasses everything from a 4-core Core i5 Kaby Lake CPU to an 18-core Core i9 Extreme Edition, which is a Skylake CPU. For clarity's sake, the Kaby Lake CPUs, also called Kaby Lake-X, are the Core i5-7640X and the Core i7-7740X. The rest of the Core i7 and Core i9 CPUs are Skylake, collectively called Skylake-X.
This union has been greeted with some confusion and trepidation. It's likely that X299 motherboards will be expensive. Some are rightly wondering who would buy a $350 motherboard to install a $250 CPU.
Intel’s motives for the Kaby Lake-X may actually be a nod to the over-clocking sports. Unlike LGA1151-socketed Kaby Lake chips, Kaby Lake-X chips have no integrated graphics capability. In fact, we’ve been told the chips physically have no IGP on the die at all. This allows the two Kaby Lake-X chips to overclock potentially far higher than the LGA1151 versions. At the recent Computex show in Taipei, in fact, Intel said a record was set for the highest overclock of a Kaby Lake chip, and it was on X299.
In a perfect world, we’d all have 18-core CPUs, but the truth is there are those who buy cheap CPUs on nice motherboards. Kaby Lake-X is for you.
We’re rationing PCIe lanes still?
Still, having Kaby Lake-X and Skylake-X on the same socket is bound to create confusion. Exhibit A is the PCIe rationing. With the Core i9-7900X, for example, you get quad-channel memory support and 44 PCIe Gen 3 lanes directly from the CPU. If you were to drop a Core i7-7740K into the same build, the motherboard drops down to dual-channel memory support and perhaps worse: the PCIe lanes drops down to 16 lanes because the Kaby Lake core doesn’t support more. That means some of the slots on a motherboard would fall back in performance or be completely disabled.
While Kaby Lake-X’s 16-lane limit is due to the CPU’s design, Intel dials back PCIe lanes on Skylake-X intentionally. Rather than the 44 lanes the 10-core version gets, Intel skimps and gives the 6-core and 8-core versions of Skylake-X 28 lanes. From what we understand, there’s no technical reason for it, just “market segmentation,” which is a business school way of saying, “so we can charge you more.” Oy.
Even more controversial than PCIe rationing is Intel’s VROC, or Virtual RAID On CPU. It’s a nifty feature on Skylake-X that allows a user to configure up to 20 NVMe PCIe drives in RAID into a single-bootable partition.
The problem? Intel apparently intends to charge more money for the feature. Although Intel still hasn’t supplied details, vendors at Computex told us they believed RAID 0 would be free, RAID 1 would cost $99, and RAID 5 and RAID 10 could cost $299. Once you ponied up the cash, you'd get a hardware dongle that would unlock the feature.
It gets worse: VROC will work only with Intel SSDs and pricier Skylake-X parts. If you buy Kaby Lake X, you’re shut out. VROC also applies only to PCIe RAID that runs directly through the CPU’s PCIe lanes. X299 still supports various RAID 0, 1, 5, 10 through the chipset, but the chipset RAID won’t touch the performance you get from VROC.
We'll reserve final judgment until Intel confirms pricing. Considering the face-palms it caused at Computex, we'll be interested to see how this shakes out.
How Core i9 changes Skylake
Once you’ve gotten beyond the platform controversy and confusion, there’s a reward. The actual Skylake-X chip itself is indeed something to admire, because it’s built unlike any previous high-end Intel consumer chip.
Previous “enthusiast” or “extreme” CPUs have mostly been the same. That is, a 4-core Haswell Core i7-4770K wasn’t all that different from from an 8-core Haswell-E Core i7-5960X except for the support of quad-channel RAM.
With Skylake-X, Intel breaks with tradition, with some major tinkering under the hood. The most noticeable is an increase in Mid-Level Cache (MLC) or L2 cache: Intel has quadrupled it to 1MB per core, up from 256MB in last year’s Broadwell-E and the majority of Intel’s CPUs. The Last Level Cache (L3) actually gets smaller, with 1.375MB per core vs the 2.5MB of the previous Broadwell-E chip, but Intel compensates with the larger MLC and also the use of a non-inclusive cache design. Compared to the inclusive design in Broadwell-E, which might keep data that's not needed, non-inclusive cache cache attempts to track what should be in the cache so it can more efficiently use the available space.
Intel also ditches the ring-bus architecture it has used for several years (including Kaby Lake and Skylake) for a new mesh architecture. If you think of a quad-core CPU as four homes connected by a bus line that makes stops at each home, it’s perfectly fine until you, add, say, 12 or 18 homes to the community. You could connect two bus lines together, but that still isn’t as fast as simply driving from one home to the next, which is what the new mesh architecture does.
Intel’s use of a mesh design clearly puts it on a better footing to compete with Threadripper, as more and more cores are added to CPUs. AMD’s Ryzen series uses something it calls an Infinity Fabric, which is essentially a super-high-speed mesh network.
The last feature worth noting is the improved Turbo Boost Max 3.0. This is the nifty feature where Intel identifies the “best” CPU core at the factory and gives it a little more boost speed. With Broadwell-E CPUs, only one core was chosen. With Skylake-X, two cores are identified as the “best” and allowed to run a couple of hundred megahertz faster.
Keep reading to get to the fun part: Benchmarks, benchmarks, and more benchmarks.
Intel sent us the Core i9-7900X in an Asus Prime X299-Deluxe motherboard. We ran the testbed with the Anniversary Update build of Windows 10. Yes, we know, the world has moved on to the Creators Update, but to maintain comparability with past CPUs we stuck with this earlier build.
All of the systems (except where noted) used a GeForce GTX 1080 Founders Edition, 32GB of DDR4/2133 and HyperX 240GB Savage SATA SSDs. For our Adobe Premiere CC 2017 test, the source project and the target drive used a Plextor M8pe PCIe SSD in all but the Core i5 and the Ryzen 5 CPUs. This exception is due to a problem with the Ryzen 5’s motherboard, which failed to recognize the Plextor drive. A Samsung 960 Pro NVMe SSD was swapped in.
Where we sourced from our previous tests, those tests used the same Nvidia drivers, the same OS and the same hardware that we used for this Core i9-7900X review. We did, however, decide to update the testbed for the original 10-core Broadwell-E Core i7-6950X. That test was originally conducted on a very early Asrock X99 motherboard that didn't fully support Intel’s new Turbo Boost Max 3.0 technology. This time around, we used the same Asus X99-Deluxe II that we used for testing the two Broadwell-E chips in the comparison pool.
Cinebench R15 Performance
Our first test is Maxon’s Cinebench R15. It’s a free benchmark based on the same rendering engine used in Maxon’s Cinema4D product. It scales well with core count and frequency and is pretty much a pure CPU test. The results speak well for the the 10-core CPUs when compared to the 8-core parts. Even though we’re increasing thread count by only 22 percent, we’re seeing almost a 30-percent increase in performance.
The difference between the 10-core Broadwell-E Core i7-6850X and the 10-core Skylake-X Core i9-7900X is less than expected. According to Intel, you might see up to a 10-percent difference in multi-threaded tasks and up to 15 percent in single-threaded tasks when compared to the Broadwell-E 10-core. In Cinebench, we’re seeing just about 3.5 percent.
What changed? The motherboard. What we’re likely seeing is a result of more than a year of tuning by Asus of its X99 platform. It just pushes the CPU far harder and far faster than the first motherboard. Our initial review of the CPU in this test gave it a score of 1,792, which is quite a bit off from the 2,107 we’re seeing from it now. Other initial reviews put the chip in the low-1,800s. If that score remained true, Skylake-X would be almost 20 percent faster than Broadwell-E.
We also ran Cinebench R15 limited to just a single thread. Because the vast majority of applications and games still rely on a single thread, the performance here is just as important as it is on multi-threaded tests. The 10-core Broadwell-E now drops back a few spots, as its clock speeds can’t keep up with the 10-core Skylake-X chip's. You can also see that the CPUs with the higher clock speeds move ahead of the 6- and 8-core chips. All except for the 10-core Core i9-7900X.
Large 8- and 10-core chips have had trouble keeping up with the spry quad-cores in high clock speeds. Intel started fixing that in Broadwell-E, but if this Cinebench result holds true, Skylake-X has the potential to hang with Kaby Lake just fine.
The second test we’ll highlight is the POV-Ray. It’s a free ray tracer that goes back to the Commodore Amiga. Using the built-in benchmark, we saw the 10-core Core i9 outpace the 10-core Core i7 by about 8 percent, which is closer to Intel’s 10-percent claim. Against the 8-core Core i7-6900K, we’re seeing Skylake-X outpace it by roughly 36 percent.
POV-Ray also has a single-threaded test. Although the Core i9-7900X can’t quite keep up with the spry Core i7-7700K, it’s pretty close. We’re seeing about a 13-percent gap between the Core i9 and the Core i7-6950X, too, which is just a hair shy of Intel’s claim of 15 percent.
Our last 3D rendering test is the open-source Blender test. Using Peter Pan’s popular BMW benchmark we’re seeing a scant 2-percent gap between the 10-core Core i9 and 10-core Core i7 chips.
For kicks, we also used Blender on AMD’s custom Ryzen workload (which you can find here.) The performance difference between Core i9 and Core i7 is minimal.
Moving onto compression tests, we used WinRAR’s built-in benchmark to measure the compression performance of the various chips. One thing you’ll notice is we no longer break out the performance of the Ryzen 5 1600X and the Core i5-7600K CPUs. That’s because both of those were tested with the RAM set at DDR4/2933. Memory bandwidth doesn't matter that much in 3D rendering tests, but it definitely can tilt the scales in compression tests. Rather than cloud the results, we’re dropping them.
One thing you’ll notice is that the 10-core Core i9 suddenly takes second place to the 10-core Core i7 chip. We also tried this test with the latest beta version of WinRAR and saw no change. We surmised this might be the cache design of the new chip, but after talking with Intel, the company suggested it could be the new mesh design.
A second compression test we ran is 7-Zip’s built-in benchmark. Like WinRAR, we suspect it prefers the cache design of the Core i7-6950X more than the new one in the Core i9-7900X. It’s not enough to matter, but the upshot is there are going to be some applications favor Broadwell-E over Skylake-X.
Adobe Premiere Creative Cloud 2017 Performance
For video editing, we tasked the CPUs with exporting a video shot by our studio on a Sony 4K camera. The project was exported using the Blu-ray preset and the Maximum Render option enabled in Premiere, which helps when video is resized. We also opted to use the Mercury software engine, which relies on the CPU for the encode rather than the GPU. Many use the GPU for encoding today, but CPU encoding is still the standard for image quality.
The Core i9-7900X registered about a 7-percent advantage over the Core i7-6950X CPU, which isn’t bad, and close to the “up to 10 percent” Intel promised.
We know people will say none of this matters because “I use my GPU for encoding,” so we also ran the same test using the Mercury CUDA engine in Premiere CC 2017. This means the GeForce GTX 1080 was tasked for much of the heavy lifting.
We saw an immediate improvement in export times, but if you look at the results—the CPU still matters. In fact, the 10-core boxes still win by a decent amount. If you were encoding a multi-hour project, the 10-cores would be worth the extra cash.
Our last encoding test uses the free Handbrake to convert a 30GB 1080p MKV file using Handbrake 0.9.9 Android Tablet preset. The test is multi-threaded and scales well with clock speed. The winner is the Core i9-7900X, which comes in—cha-ching—about 10 percent faster than the 10-core Core i7-6950X. We’re also seeing nice scaling: The 10-core is about 30 percent faster than the 8-core Core i7-6900K and 60 percent faster than the 6-core Core 7-6800K.
3DMark Fire Strike Performance
For gaming performance, we first run Futuremark’s 3DMark Fire Strike. We’re reporting only the physics portion of the test, as that’s the only one that matters for the CPU. The test uses a real-world physics engine that scales well with core count. Oddly, the Core i7-6950X nudges the Core i9-7900X out of the way, perhaps because of the cache difference between the chips or the mesh architecture. Note, though, that this is a theoretical test of what a game could do it if stressed all those cores. In reality, games don’t devote this much to game physics.
Tomb Raider Performance
Moving on to a real game, we use Ubisoft’s older Tomb Raider to measure CPU performance by running the game at 19x10 resolution and the normal preset. At this low game setting and relatively low resolution for a GeForce GTX 1080, the only difference we’re likely seeing is clock speed. Each CPU’s cache can occasionally move the needle, too.
The Ryzen 7 1800X chip performance is off, likely due to code that isn’t optimized for its micro-architecture. Case in point, Rise of the Tomb Raider recently received an update that greatly helped Ryzen out. And for the most part, it’s not an issue at higher game settings where the GPU is the bottleneck on performance.
You can also see from our results why Ryzen’s performance was so confusing: CPUs don’t matter in conventional gaming as much as people wish they would.
Tom Clancy’s Rainbow Six Siege Performance
We’ll close out our gaming performance of Skylake-X with the more modern Tom Clancy’s Rainbow Six Siege. The Core i9-7900X is slightly slower than the Kaby Lake and 8- and 10-core Broadwell-E chips. We suspect, again, that the game slightly favors the cache design of Broadwell-E and Kaby Lake, but it’s not a big deal. Lackluster Ryzen performance is again possibly linked to game optimization.
One difficult thing to discern with most of these benchmarks is just how “efficient” each CPU micro-architecture is. One way to coax that out is by running a test using a single thread, with all the CPUs locked to the same clock speed. We locked most of the CPUs here to 2.5GHz and turned off any Turbo Boost. We then ran Cinebench R15.037 (which we used to generate scores for the older CPUs).
We can see that instructions per clock (IPC) has built itself up slowly from the days of Sandy Bridge. Skylake-X comes out in front of even Kaby Lake, surprisingly.
Note: The FX-8370 is fast because we couldn’t underclock the CPU to 2.5GHz, so we just used its score running at 4GHz. Yes, the performance of Vishera at 4GHz is still below that of a Haswell CPU running at 2.5GHz. Woof.
IPC isn’t everything, so we’ll close this out with a big chart of Cinebench R15 scores that we’ve personally run on various CPU models. Keep in mind, they’re not exactly one-to-one, as some CPUs have higher-clocked RAM, or run on DDR3 instead of DDR4. But Cinebench is mostly a CPU benchmark, so memory bandwidth doesn’t affect it as much as it would in some tests.
Core i9 has the performance, but the price?
If you point your eyes at that last chart, with just about every Intel design represented since Sandy Bridge, you can only conclude that the new Core i9-7900X is indeed the fastest consumer CPU ever produced by Intel. There’s just no argument. The fact that it’s being offered at $1,000 compared to the $1,723 tag on the previous 10-core is another reason for the PC community to cheer.
The problem is, it’s a different world now. At $1,000 for a 10-core chip, you’re paying about 100 percent over an 8-core Ryzen 7 chip for about 30 percent more performance. Even worse, we still don't know what price AMD set for its 12-core and 16-core Threadripper chips. If AMD introduces a 12-core CPU at $850, as some think, a 10-core Core i9 for $1,000 loses its appeal.
For now, the Core i9-7900X reigns as the fastest consumer CPU on the planet. But it should be looking over its shoulder, as will we, for Threadripper.