The new X299 platform replaces X99, and scales both lower and higher.

AMD announced its new high-end desktop (HEDT) platform, the 16-core 32-thread ThreadRipper a couple of weeks ago. At Computex in Taipei, it is Intel's turn to update its HEDT platform, and it is one-upping AMD in the process. The Intel platform, consisting of the new X299 chipset and new X-series processors, will go all the way up to 18 cores and 36 threads.

The HEDT segment is aimed at gamers, video streamers, and content creators with deep pockets or an insatiable desire for more concurrent threads than the mainstream processor segment has to offer. The value proposition for this segment is always a little skewed, with the chips being as much prestige parts as anything else. Straightforward gaming workloads may struggled to make full use of the chips' resources, but serious Twitch streamers, for example, can make good use of the extra cores. Software developers are another group that can make good use of all those cores.

The new platform has wider range and greater complexity than the X99 platform it replaces. At the low end are "Kaby Lake-X" processors: the i5-7640X and the i7-7740X. These chips are very similar to the existing mainstream Kaby Lake processors that came to market earlier this year; four cores with either four (for the i5) or eight (for the i7) threads, two memory channels, and 16 PCIe 3 lanes from the CPU itself. The chips have a higher power envelope—up to 112W, instead of the 91W of non-X parts—and use X299's new Socket 2066. Their clock speeds are a little higher, too; the $339 i7-7740X has a base clock of 4.3GHz and a turbo of 4.5GHz, compared to the 4.2/4.5GHz of the $339 i7-7700K.

Step up the price a little and things become more interesting. Out goes Kaby Lake-X, in comes Skylake-X. Skylake-X is not simply an existing Skylake processor in a new socket, instead, it's a desktop version of the Skylake-SP core that's going to be used in the next generation of Xeon chips. Skylake-X will include some of the new features of Skylake-SP, such as the new AVX-512 instruction set that both doubles the size of the vector operations (up from 256-bit) and includes new instructions to make those vector operations more flexible. However, Skylake-X will also retain certain design decisions found in the X99 platform: while Skylake-SP will support up to six memory channels, for example, X will be limited to four.

Skylake-X also introduces the latest revision of the Turbo Boost Max 3 feature that was used on the previous-generation Broadwell-E HEDT chips. In Broadwell-E, Turbo Boost Max 3 identifies which individual core within the processor supports the highest clock speeds and boosts that core in workloads that stress only a single core. Skylake-X extends that to identify the two individual cores that support the highest speeds, going from a single-core boost to a dual-core boost.

Skylake-X also shakes up the cache topology used in the new chips. It's neither the same as the hierarchy used on existing Skylake chips, nor is it the same as was used in Broadwell-E. Each individual core now gets a 1MB private level 2 cache, up from 256KB in Broadwell-E and Skylake. The shared last-level cache, however, has been shrunk; instead of the 2.5MB per core of Broadwell-E, or the 2MB/core of Skylake, it's now only 1.375MB/core. That last level cache has also been changed from an inclusive design (wherein it contains everything held in the per-core level 1 and level 2 caches) to a non-inclusive design, wherein it only contains data that has been evicted from the private caches.

Intel says this switch from large shared caches to a smaller shared cache and large private caches boosts performance; the private caches have lower latency than the shared cache, and the quadrupling in size means they have a superior hit rate to the previous generation caches. There are sure to be trade-offs though; the old inclusive design meant that one core could read data held in another core's cache merely by accessing the last level cache. To do the same now would require the first core to snoop the level 2 cache of the second core.

The Skylake-X chips will also expand Intel's numbering system: Intel is adding a new i9 branding that slots in above the i7 branding for the high-end processors.

As is Intel's wont, the Skylake X range will suffer certain kinds of segmentation. The cheapest Skylake X part, the $389 six core, 12 thread i7-7800X, won't include Turbo Boost Max 3 and will only officially support 2400MHz memory. Both that processor and the next highest, the $599 eight-core, 16-thread i7-7820X, will have only 28 PCIe 3 lanes, though this part will support memory speeds up to 2666MHz. It's not until the $999 10-core, 20-thread i9-7900X that the full range of features is lit up: 44 PCIe lanes, Turbo Boost Max 3, and 2666MHz memory. These three parts will all have a 140W power envelope.

Beyond that, and with a 165W power rating, parts will be available with 12, 14, 16, and 18 cores (and twice as many threads), with that top 18-core, 36-thread chip, the i9-7980XE (branded not merely i9 but "i9 Extreme" on the box), costing a whopping $1,999. Intel hasn't yet disclosed clock speeds for these parts.

All the chips will include unlocked multipliers, making them overclocking-friendly. The X299 chipset also supports Intel Optane Memory for disk acceleration, up to three PCIe/NVMe SSDs, 8 SATA ports, and 10 USB 3.1 generation 1 ports. While the X299 chipset will be supported when used in conjunction with an Alpine Ridge Thunderbolt 3 and USB 3.1 generation 2 controller, the high-speed interconnect won't be built in. That's going to be reserved for some future processor generation.

The inclusion of Kaby Lake-X processors, at prices that match their mainstream editions, gives the new X299 platform a little more reach than the old X99 platform. Buyers who are considering an i7-7700K paired with the Z270 chipset might instead plump for an i7-7740X with X299. The motherboard will, in all likelihood, be a bit more expensive, but the higher power rating, higher base clock speed, and omission of the integrated GPU should give a bit more headroom for overclocking. The ability to drop in a Skylake-X will also give the cash-rich gamer a solid upgrade path should games start to take more consistent advantage of high thread count processors.

The new platform also shows that Intel isn't willing to cede territory to AMD. Intel's chips retain their price premium; the eight-core, 16-thread Intel chip costs $599, whereas AMD's corresponding part is $499. The Intel chip does have twice as many memory channels, so it's not exactly like-for-like, but we'd expect that you'll be paying at least a little bit more for an Intel processor and an Intel motherboard than a roughly similar AMD system. But with its 18-core chip, Intel is scaling a bit higher for those who truly want the best that money can buy. Intel's single-threaded performance, which is still critical in many workloads, is also going to be superior; AMD's Zen slightly trails Broadwell-E in single-threaded performance and operates at a more significant disadvantage relative to Skylake and Kaby Lake.

X299 motherboards and X-series processors should hit the market "in coming weeks."