AMD’s Ryzen CPUs and their oodles of cores and threads have taken the market by storm of late, and the company seems poised to deliver even more computing power this summer with the Ryzen Threadripper high-end desktop chip. Of course, you can’t make noise like that in the CPU market without awakening a sleeping giant.
Intel is reasserting its x86 dominance this morning with a broadside of Core X-series high-end desktop CPUs, and it’s also introducing a new host platform with the X299 chipset and LGA 2066 socket. This platform can host chips ranging from four cores and four threads all the way up to a monster 18-core, 36-thread Extreme Edition chip: the Core i9-7980XE.
Before we talk about the summit of the Core X-series lineup, let’s look at base camp. In what may be a first for Intel’s high-end desktop chips, the Core X-series counts two different generations of microarchitectures among its members. At the base of the pyramid, the four-core, four-thread Core i5-7640X and the four-core, eight-thread Core i7-7740X transplant Intel’s highest-performance consumer Kaby Lake chips onto the X299 platform with higher TDPs, higher base clocks, and dual-channel DDR4-2666 memory support.
Intel seems to think that enthusiasts will choose these Kaby Lake-X chips as a gateway to future upgrades on their X299 mobos over their LGA 1151 counterparts. That might be an interesting prospect for some enthusiasts, but aside from their higher base clocks and TDPs versus their LGA 1151 doppelgangers, Kaby Lake-X CPUs aren’t the most intriguing parts Intel is introducing today.
The Core i7-7800X on up is where the action really starts to happen for this launch. Intel has tapped its Skylake Xeon production line to repurpose server silicon for X299. The Skylake-X dies (likely 18-core and 12-core base dies) that Intel is using for these parts gives the company a wealth of knobs and dials to tweak as it creates products, and tweak it has. The Skylake-X family boasts a whopping seven chips this time around, compared to the usual four Intel has historically launched with each high-end desktop refresh.
Unfortunately, Intel played its cards rather close to the chest about the highest-end Skylake-X CPUs ahead of today’s Computex keynote. We do know that starting with the 10-core, 20-thread Core i9-7900X, Intel will be using the Core i9 brand to set apart its most extreme desktop processors. The 12-core, 24-thread Core i9-7920X, the 14-core, 28-thread Core i9-7940X, the 16-core, 32-thread Core i9-7960X, and the beastly 18-core, 36-thread Core i9-7980XE (for Extreme Edition) all herald a new era of computing horsepower for the high-end desktop PC— all for a pretty penny, of course. More on that in a second.
The 18-core Skylake-X die. Source: Intel
Aside from those core counts, however, we know next to nothing about Core i9 CPUs as of this writing. Base clocks, Turbo clocks, and Turbo Boost Max 3.0 speeds all remain under wraps as of Intel’s Computex keynote. We do know that those many-core Core i9s will not have modest thermal envelopes. Expect a 165W TDP from the Core i9-7920X on up, and expect clock speeds to decrease as core counts climb.
Intel strongly recommends liquid cooling for these chips, and it’ll be offering an Asetek-produced 120-mm all-in-one cooler for LGA 2066 CPUs called the TS13X. Folks looking to overclock Core i9 CPUs will likely need even more substantial cooling hardware to extract the full potential from their chips.
Even with this dearth of basic information, the architectural improvements in Skylake-X might give AMD’s Ryzen Threadripper CPUs reason to look over their shoulders as the red team prepares those many-core parts for launch later this year. Let’s delve into why now.
Skylake-X gets wider and cachier
While we don’t have full details of how the server version of Skylake stacks up against its mainstream desktop counterpart yet, Intel has let slip a few details. First off, Intel has implemented AVX-512 support with Skylake-X. That means the width of each AVX data register has doubled compared to the desktop version of Skylake, which uses 256-bit-wide registers.
Source: GNU GCC
AVX-512 support also nets 32 data registers for Skylake-X, compared to 16 for typical desktop Skylake. In theory, AVX-512 will yield a doubling of the FLOPS per second per core compared to Intel’s older architectures. You’ll hear Intel call the Core i9-7980XE its “first teraflop CPU,” and the move to AVX-512 is why. Assuming developers recompile their software to use AVX-512 instructions, SIMD-heavy programs could get a big speedup from these chips.
In perhaps the biggest change to its microarchitecture, Skylake-X also introduces a new cache hierarchy. Each core now has a 1MB private L2 cache, up from 256KB in mainstream Skylake cores. The shared last-level cache, or L3, for Skylake-X has shrunken a bit in turn, to as much as 1.375MB per core. Broadwell-E boasted as much as 2.5MB of L3 cache per core. This new L3 structure is no longer inclusive of the L2 cache, either, as it was with Broadwell-E. As a rule of thumb, quadrupling the cache size as Intel has done here will cut the miss rate in half, assuming the same associativity. We’ll be digging for more details of these changes when we meet with Intel about its X-series CPUs soon, but expect a major performance boost from this rejiggering of the caches on Skylake-X regardless.
It’s not strictly a microarchitectural feature, but Skylake-X also debuts a new take on Intel’s Turbo Boost Max 3.0 technology, first seen with Broadwell-E chips. In the first iteration of Turbo Boost Max 3.0, Intel identified the highest-boosting core on a Broadwell-E die. In tandem with a Windows driver, TBM 3.0 could let those high-potential cores boost well above their normal Turbo speeds for high performance in lightly-threaded workloads. This time around, Intel is snooping out the best two such cores on a Skylake-X die for even more of a performance improvement, and no more Windows driver is needed to let TBM 3.0 work its magic. Intel also tells us to expect more consistent support and implementation of the TBM 3.0 feature from motherboard manufacturers this time around, a welcome improvement compared to the state of the feature on Broadwell-E motherboards.
The X299 platform takes a dip in Kaby Lake
To host Skylake-X and Kaby Lake-X CPUs, Intel is introducing a new socket, called LGA 2066, and a new chipset, called X299. We’ll be covering many, many X299 motherboards this week from Computex, but here are the basics. LGA 2066 uses more pins than LGA 2011v3 before it, but its outer dimensions are identical to the older socket’s. Builders will obviously need new motherboards to build with Core X-series chips, but they’ll at least be able to reuse their existing cooling hardware.
The X299 chipset might also evoke some deja vu among observant gerbils. That’s because Intel based the X299 PCH on its consumer Kaby Lake chipsets, rather than following in the footsteps of the server-sourced X99 PCH. The infusion of Z270 in the X299 chipset offers several benefits to Intel’s high-end desktop platform, although it breaks little new ground. That said, X299 upgrades the link between the CPU and the chipset to DMI 3.0, which offers roughly the same bandwidth as four lanes of PCIe 3.0. Broadwell-E still relied on the X99 platform’s DMI 2.0 link.
X299 also gives motherboard makers 30 flex I/O lanes to play with. Those flex lanes can be configured as up to 24 PCIe 3.0 lanes, as many as eight SATA 6Gbps ports, and up to 10 USB 3.0 ports. Thanks to its Z270 roots, however, X299 still doesn’t boast onboard USB 3.1 Gen2 support, so expect that connectivity to come by way of external controllers on X299 mobos.
X299 will also support Intel’s Optane Memory and Optane SSD technologies. We doubt many X299 builders will still be considering using hard drives as boot devices, but Intel plans to offer a full-bore 3D Xpoint SSD later this year that seems like a perfect complement for a Skylake-X system. When that storage device arrives, X299 PCs will be ready for it.
Pricing and positioning
Many hoped (myself among them) that Ryzen’s arrival would force Intel high-end desktop CPU prices down compared to their historical stickers, and for two chips in the Skylake-X lineup, that’s sort of what’s happened. The Core i7-7820X brings eight cores and 16 threads to the $600 slot for the first time in Intel’s history, while the Core i9-7900X shifts the 10 cores and 20 threads that were previously the domain of the $1650 Core i7-6950X down to a grand. I was disappointed that Intel didn’t choose to add more cores for the same prices as Haswell-E chips when it launched Broadwell-E, so it’s heartening to see the company returning to that practice with Skylake-X.
Meanwhile, the six-core Core i7-7800X rings in at the same $389 price tag as the Core i7-5820K of Haswell-E fame. Compare that to the $434 one had to pay to get into the six-core Core i7-6800K, and it seems Ryzen CPUs are keeping Intel honest at the low end of the high end, if nothing else.
Of course, Intel couldn’t resist taking away as it gave us more cores for less money. To get 44 lanes of PCIe connectivity from the CPU with Skylake-X, one now has to pay $1000 for the Core i7-7900X. Both the Core i7-7800X and Core i7-7820X have the dreaded 28 lanes of connectivity that used to be the sole domain of the most entry-level Intel HEDT CPUs, a move that will doubtless leave some builders a little sour as they plop down $600 for an i7-7820X.
Even though Intel’s decision to make the full 44 lanes of PCIe 3.0 more exclusive stings here, it might not sour builders as much as it might have in the past. Interest in multi-GPU as a route to higher graphics performance seems to be dwindling by the day, and multi-GPU was the primary reason we dinged Intel’s 28-lane i7-5820K when it first appeared. Times have changed since. A single GTX 1080 Ti is enough graphics card to power all but the most excessive gaming systems, even with a 4K screen hooked up. Assuming one decides to chance SLI’s hit-or-miss game support with twin GTX 1080 Tis, we’re guessing they can afford a $1000 CPU or better to go with them.
The biggest demand for PCIe these days beyond graphics is for storage devices, and the lower-end Skylake-X parts still do well enough here. Motherboard makers could conceivably wire up three full-speed M.2 slots or U.2 ports from the remaining 12 lanes on the i7-7800X and i7-7820X, and that’s before we consider what’s possible with 44 lanes from the higher-end chips. Mobo engineers should also have fun with the freedom of X299’s flex I/O lanes, although devices connected to those lanes will be bottlenecked by the roughly PCIe 3.0 x4 bandwidth of the chipset’s DMI 3.0 link.
What about Core i9 CPUs? Well, more cores mean more money, but the Core i9 lineup actually seems pretty reasonably priced for what it offers, assuming you need more cores. The Core i9-7920X’s 12 cores are a $200 upcharge over the $999 Core i9 7900X’s 10-core, 20-thread complement, and another $200 jump gets you 14 cores and 28 threads on board the Core i9-7940X. From there, the delta changes to $300: first for the $1699 Core i9-7960X, and then to $1999 for the Core i9-7980XE big daddy with its 18 cores and 36 threads.
Even at $1999, the Core i9-7980XE seems like a fair bargain compared to many-core Broadwell Xeons. As just one example (and presuming your X99 motherboard offered support for it), one could get 18 Broadwell cores and 36 threads in the Xeon E5-2695 v4 for $2400 at retail. X299 likely won’t offer server-grade niceties like ECC RAM support, but there’s no denying that it’s making a certain class of compute capability more accessible—and with higher potential performance—than ever before on Intel’s high-end desktop platform.
Intel says Core X-series CPUs will be available “in the coming weeks,” and we’ll be learning more about them soon. For now, however, you can join us in pinching ourselves about the prospect of 18 Skylake-X cores in a high-end desktop PC. Hopefully we can stop dreaming and start testing something like the Core i9-7980XE soon.