Qualcomm's 14nm "Taipan" Core: What We Can Expect From The Krait Successor

Qualcomm's 2015 roadmap has apparently leaked, and it includes information we've never seen before, such as the existence of a "TS2" core that could be Qualcomm's next-generation custom core and the successor to the "Krait" architecture. The leak also points out that Qualcomm will be using Samsung/GF's 14nm FinFET process this year, although only for its high-end Snapdragon 820 SoC.

Qualcomm launched the Krait custom core back in the spring of 2012, so the new TS2 core, likely based on the "Taipan" architecture, would arrive three and a half years later. That is almost twice as much as one of ARM's, or even Apple's, two-year long generations. Being in a dominant position in the mobile chip market has allowed Qualcomm to squeeze as much time out of the Krait core as possible before it becomes obsolete.

In late 2013, Apple launched the "first mobile 64-bit core," surprising almost everyone -- including Qualcomm, which didn't expect chips based on ARM's new ARMv8 instruction set to come out until at least late 2014, if not later.

There have been hints that Qualcomm was not prepared for Apple to launch a 64-bit chip. This could have forced Qualcomm to make two decisions: 1) to come out with a "64-bit chip" as soon as humanly possible (how we got the Cortex A57-based Snapdragon 810) and 2) to finish, and possibly overhaul its Krait successor, based on what the company then knew about Apple's 64-bit core.

The fact that Qualcomm's Snapdragon 410 would be based on Cortex A53 didn't surprise many when it was announced. Qualcomm has already been using Cortex A7 (the predecessor to Cortex A53) in the Snapdragon 400 as a way to cut costs and compete against the likes of Mediatek and other Chinese SoC makers.

Cost is a bigger factor at the lower levels of the mobile market than differentiation (although that matters as well). However, at the high-end, differentiation is critical, as consumers there tend to be more tech-savvy about comparing the chips being put into their smartphones.

The lack of differentiation (at least on the CPU side) in the Snapdragon 810 did surprise many. Qualcomm hasn't used an ARM stock core since the whole touchscreen smartphone market started taking off. Even in the early days it had the custom Scorpion core, predating Krait.

This information, coupled with the new leak that shows Qualcomm is going to replace Cortex A57 at the high-end not even half a year later, seems to prove that Qualcomm did indeed try to get a 64-bit chip out of the door as soon as possible.

As for Qualcomm's second decision, to overhaul its Krait successor, things are not as clear. If the company didn't overhaul it, then Qualcomm has a rather long (three and a half year) mobile architecture design cycle, compared to its competition.

If Qualcomm overhauled the Taipan design after the company saw Apple's Cyclone core, then it probably didn't change all that much. Instead of going with wide two-core CPUs as Apple did, Qualcomm has chosen to go with an eight-core design, much like with the big.LITTLE-based Snapdragon 810. In Snapdragon 810's case, four of those cores are low-end and four are high-end. The Snapdragon 820 seems to have eight equally high-end cores.

However, those cores are likely not as wide as Apple's Cyclone or Nvidia's Denver cores. The "Taipan" core may very well be mostly a natural evolution of the Krait core that's both higher-performance and more efficient.

ARM has said that big.LITTLE is a better CPU design because it puts some low-power cores in charge of low-end tasks, and some high-performance cores in charge of complex tasks. It will be interesting to see what kind of case Qualcomm will make for having eight very similar cores, being in charge of all tasks, from the simplest to the most complex.

Are they actually being utilized by the Android operating system most of the time, or just very rarely? If only rarely, then what are those types of applications or situations that need the full eight cores? Are they important enough to require the extra cost of four or six more cores? How much better is having eight cores than two or four wider cores, even for those special applications?

These are all questions Qualcomm should be answering when it launches the new "Taipan" core.

Not unexpected, but still good to see, is the fact that the Snapdragon 820, Qualcomm's flagship chip for fall 2015, will be built on 14nm FinFET, similarly to Samsung's own Exynos chips and Apple's A9 chip. Qualcomm has been late to the 20nm party. The 20nm Snapdragon 810 is still not out yet, while Samsung's Exynos 5430, Exynos 5433, Exynos 7 Octa and Apple's A8 have been on 20nm since the second half of last year.

One of Qualcomm's key success factors in 2012, when it launched Krait, was the adoption of the cutting edge (at the time) 28nm process node. Even if it's more expensive than using a slightly older node, adopting the most cutting edge process seems to almost always pay off. Both the OEM customers and the end-users of the devices want the absolute best at the high-end and are willing to pay extra to get it.

Nvidia, on the other hand, has almost always gotten this strategy wrong. It usually prefers to go with cheaper but older process nodes, making Qualcomm's Snapdragon an easy choice for most OEMs and consumers.

Qualcomm will also have a few more SoCs ready by the end of the year, including the Snapdragon 616, a higher-clocked Cortex A53-based successor to the Snapdragon 615; the Snapdragon 620, a Taipan-based successor to the Snapdragon 610; and two more octa-core chips built on 20nm, the Snapdragon 625 and 629. Qualcomm will also release a more direct successor to the Snapdragon 810 called the Snapdragon 815, which will be based on the same big.LITTLE design as the 810.

Although this is supposedly Qualcomm's roadmap for 2015, it's possible that some of these chips won't make it to market this year. The one with the highest chance for that happening is actually the hero SoC here, the Snapdragon 820. Its launch will likely depend on how soon or late the Snapdragon 810 arrives on the market.

At the same time, Qualcomm will probably want to avoid missing out on the Galaxy Note 5 after presumably already being left out of 90 percent of Galaxy S6 devices. In that case, the company will probably want to show its customers the best chip it can make at the time.

Follow us @tomshardware, on Facebook and on Google+.

Lucian Armasu
Lucian Armasu is a Contributing Writer for Tom's Hardware US. He covers software news and the issues surrounding privacy and security.
  • wishmaster12
    FinFet looks interesting, if they can put more of them on one chip and get a high clock frequency without frying the chip!
  • InvalidError
    15105742 said:
    FinFet looks interesting, if they can put more of them on one chip and get a high clock frequency without frying the chip!
    Frying the chip is not going to be an issue since they will be only squeezing as much extra performance as they can within the same or preferably smaller thermal/power profile - a more power-efficient SoC means more power budget left for extra screen, talk or standby time on a given battery.
  • aldaia
    15105742 said:
    FinFet looks interesting, if they can put more of them on one chip and get a high clock frequency without frying the chip!

    Finfet has been around for quite some time. It was first proposed in 1999 by Berkley researchers. "Huang, X. et al. (1999) "Sub 50-nm FinFET: PMOS" International Electron Devices Meeting Technical Digest, p. 67. December 5–8, 1999" And later implemented by Intel in late 2011 under the name tri-gate or 3D transistors.