Intel has just unveiled its next-generation Atom CPU core. Called Silvermont, this 22nm CPU will power the Merrifield (smartphone) and Bay Trail (tablet) SoCs. While its predecessor, Saltwell (Medfield & Clover Trail), was the first Intel chip to earn a place at the mobile table, Silvermont is the first product released by Intel that should take the performance and performance-per-watt crowns from ARM designs made by Qualcomm, Nvidia, Apple, and Samsung.
With the PC market waning, and investors flocking to competitors with proven mobile designs, Silvermont is a very, very big deal for Intel. Is a fast mobile x86 CPU enough to save a floundering Windows tablet market, though, or are there other endemic issues at play?
Silvermont is Intel’s first new Atom microarchitecture in five years, and it has been re-architected from the ground up to provide excellent performance, while sipping fewer Watts than its ARM-based cousins.
One of the most significant changes was the shift to 22nm FinFET, which really blows the doors off any competing process technologies as far as power consumption goes. Combined with the shift to 3D transistors, Silvermont’s architecture and design were actually co-optimised with the process. The ability to design a specific process for a specific CPU core is one of the main reasons that Intel holds such a dominant position over companies such as Qualcomm and Nvidia, who have to work with TSMC’s one-size-fits-all 28nm processes.
Beyond the new process, Silvermont has a slew of new features that should improve power efficiency and performance over the Saltwell core. There’s out-of-order execution (OoOE), which allows for instructions to be executed as soon as there’s data to be processed, rather than in the exact order laid out by the software program. There’s more efficient branch processing, more accurate branch predictors, and faster recovery from pipeline crashes/collisions.
In terms of additional instructions, Silvermont borrows a lot of features from Westmere (the die shrink of Nehalem). There’s AES-NI (hardware-level encryption/decryption), Secure Key (random number generation), and SSE4.1 and 4.2. The following image shows Saltwell versus Silvermont pipeline changes (the change from 10 to 13-cycle mispredict is very nice):
One of the biggest performance boosts is likely to stem from massively improved FPU latency and throughput: Silvermont executes many instructions in half the time, while throughput has almost doubled. If you ever lamented the x87 FPU performance of previous Atom cores, be cheered: Silvermont’s FPU performance should be comparatively beastly.
Rounding out the architectural changes, there’s also some high bandwidth/low latency caches and out-of-order memory transactions. For a complete rundown of the new microarchitecture, you can attempt to interpret the core block diagram below (and you’ll find an annotated version of the block diagram immediately underneath that). In short, though, Silvermont gets more done per clock cycle (IPC), and it uses much less power to perform each clock cycle. While no one actually has any Silvermont hardware to independently benchmark, Intel is claiming a ~3x performance improvement at ~5x lower power, over current Saltwell (Medfield/Clover Trail) Atom chips.
Moving from the micro to the macro scale, Silvermont finally brings real multi-core support to Atom. Somewhat like AMD’s Bulldozer, Silvermont comes in modules of two cores, which share a tightly coupled up-to-1MB L2 cache. An Atom SoC can contain up to four modules, for a total of eight Silvermont cores. There’s per-core frequency and power management – but there’s no Hyper-Threading.
On the performance front, Burst Mode (Atom’s version of Turbo Boost) has been souped up: Whereas Atom used to just base its P-states (SpeedStep) on core temperature, Silvermont manages Burst frequency by analysing thermal, electrical, and power delivery constraints. Silvermont can also dynamically share/allocate power between CPU cores and the GPU, which is obviously rather neat for mixed workloads.
The end result of all this work is best summarised with an almost unbelievable graph showing Intel’s Silvermont core versus some dual and quad-core ARM competitors:
What you don’t see here is the small print at the bottom of each slide, warning us that Intel may have used benchmarks that are “optimised for performance only on Intel microprocessors.” Caveats and loaded dice aside, though, it does seem like Silvermont will be very frugal in the power consumption department. It will be very exciting when Intel actually starts shipping dev kits that we can independently benchmark.
Mobile performance: The ultimate oxymoron
We’ve been toying with one particularly vexing question for the last couple of years: Does mobile performance really matter? Of course you want enough performance to perform your daily tasks without interface lag or load times, but the necessity for higher performance very quickly tails off after that. Qualcomm (or Intel or Nvidia) might release a chip that’s two times faster than the competition, but does that really matter? Does a faster CPU improve download speeds? Or call quality? Or battery life? Or the availability of good apps and games?
For much the same reason that we stopped comparing (and caring about) PC CPU clock speeds, the actual significance of Intel’s Silvermont performance claims are questionable. A far more important metric is power consumption, performance-per-watt, and battery life – and yet Intel gives almost no guidance in any of these areas. The previous graph certainly looks good, but Intel stops short of providing actual power consumption figures that we can compare against ARM designs. Basically, we know that Silvermont is much more powerful than Saltwell at a given TDP – but Intel won’t tell us what the TDP is.
Which brings us neatly onto the stillbirth of Windows 8 tablets. Silvermont isn’t going to magically conjure up a Metro app ecosystem. Silvermont will improve the performance of Desktop apps, but have you tried using Desktop apps on a touchscreen tablet?
You could argue that Silvermont (Bay Trail) will allow for a tablet with Surface Pro-like performance, with Surface RT-like battery life – but that’s naive, and ignores other issues, such as cost. If Microsoft couldn’t get the Surface RT below £400 with an ARM processor inside, there’s no way it’ll manage it with Silvermont. In short, a faster x86 processor isn’t going to save Windows 8 tablets.
On the other hand, if Silvermont’s performance-per-watt is really as good as Intel says, Bay Trail could enable some seriously sexy Android tablets. The only problem with that idea, though, is that Intel has sworn blind that it isn’t interested in Android tablets.
Last year, Intel’s mobile chief told me in an interview: “Windows 8 on tablets, Android on smartphones. Right now, I have as many people working on Windows 8 tablets as I have on Android phones… the way for us to succeed is to focus – through close collaboration with Google on Android, and Microsoft on Windows 8.”
Who knows whether Intel is actually telling the truth, though, or whether that’s just political posturing to keep the Wintel alliance healthy. We would be very surprised if Intel didn’t have some x86 Android tablets in its R&D labs.
Silvermont could be very exciting in the smartphone space, however. While Medfield wasn’t the fastest SoC on the block, it was fast enough – and as we’ve already noted, performance isn’t everything. Silvermont will be much faster than Medfield, while consuming less power. At this point we can’t say if Silvermont will beat the latest Qualcomm Snapdragon or Apple A-series SoCs, but it will almost certainly be fast enough to power a flagship smartphone, while hopefully not breaking the battery bank.
With Intel’s continued work on x86 Android, it now seems fairly certain that this will be the year that an Intel-powered smartphone finally arrives in the US market. If that phone is a success, Intel will finally have a mobile beachhead – and it will be then, and only then, that ARM will finally experience the big guns of Intel’s manufacturing mastery and promise to release a new Atom core every 12 months.