Intel Silvermont Architecture Updates Atom for Phones and Tablets
Power, Performance and Closing Thoughts
Intel obviously believes that one of the biggest advantages it has in the SoC market is its process and manufacturing technology. With the current tri-gate 22nm transistors they definitely have the edge and they are working hard to tweak another version of the technology specifically for low-power and SoC designs.
Some parts (like the CPU cores) require faster switching, low voltage units. I/O requires slower switching with high voltage (handling up to 3.3v in some cases). This idea is nothing new, as NVIDIA did it with Tegra and AMD did it with Brazos (different transistor types for CPU and GPU portions). However, the granularity that Intel is exposing seems to overshadow previous process nodes.
The goal as well is to include third-party IP in these designs with the Atom architecture including other graphics modules, radios, etc. By enabling more flexibility in the transistor capability of the 22nm technology Intel can offer better performing parts, analog compatibility and mixed signal RF.
Previous generations of Atom cores used boost technology but only exposed additional P-states to the CPU based on available thermal headroom. With Silvermont the x86 cores run at burst frequencies managed by hardware measuring thermal, electrical and power delivery constraints. The SoC is able to share power between the CPU cores and graphics or other included IP though it only appears to work in one direction - the CPU can borrow excess power from other IP. Allowing sharing in both directions would require a lot of additional work for the external IP and would likely only be done in some very explicit designs.
The burst operating points can be adjusted dynamically based on thermal properties of the device (larger heatsinks, etc) but there is very little overlap between what Silvermont is doing and what the Core series algorithms are doing with Turbo Boost.
Silvermont offers the capability to put each individual core into the C6 sleep state, a feature that previous designs lacked. From a module view, depending on the implementation you can have a single shared voltage rail to all modules or one rail for each to better control low power states.
One of the biggest competitors in the mobile field from ARM is the idea of the "Big+Little" designs. These processors combine larger, higher performing chips with smaller, more power efficient options and switches between them seamlessly to the user. NVIDIA implements a version of this on their Tegra 3 design with the 4+1 implementation. Intel claims though that their options are much less efficient than they at first sound with complex switching algorithms to move data from one CPU to the other and long switching times that cost efficiency.
The graph above pits Intel's claims against their own competition. Silvermont will apparently have the ability to run at continuously lower power levels while offering better performance along the way. We are still months from having our own devices to test against these claims though.
For our first graph on performance Intel is comparing Silvermont to the existing Saltwell designs. In the single core, single thread metrics we see a 2x performance jump with 4.7x lower power. With the 50% IPC increase claimed on the previous page we can tell you that we are definitely seeing higher bust clocks with Silvermont than Saltwell in the same power envelopes.
The second data set compares a dual-core HyperThreaded Saltwell design with the dual-module four core design of Silvermont. Performance increases go up to 2.5x and power efficiency is still about 4.4x better. This shows the advantages in performance of "real" cores rather than HyperThreaded in-order designs and also indicates that not much is lost on the power efficient side of things.
Intel's next claim is that inside the standard phone power target, Silvermont will offer better performance than either of the competitions dual-core or quad-core offerings.
This graph shows us a dual-core Silvermont design going up against a quad-core phone solution, measured at 1 watt core power. While they weren't willing to name names here, Intel claims that they will be anywhere from 1.4x to 2.1x faster at the same power levels OR 1.6x to 3.1x more power efficient at the competition's 1 watt core performance.
Our second performance graph shows quad-core Silvermont against quad-core ARM designs for the tablet market. Using the same style of performance comparisons, Intel claims Silvermont will be as much as 2.3x faster than any other current options at a 1.5 watt core power level OR as much as 5.8x more power efficient at the competitors PEAK performance.
For quite some time now I have been hearing about Silvermont and its chip derivatives like Bay Trail and Merrifield, and that this architectural shift would be the one that finally put Intel on the map for tablets and cell phones. From internal sources, even as Medfield was being pushed out around the world and Clovertrail tablets found their way into our labs, the silent message was always that Silvermont would be the real game changer.
After hearing Intel's engineers and product marketers talk about it, I am cautiously optimistic about what Intel has created here. The one problem I see with that data that was presented is quite simply the time delay on product release. Tablets will likely be available in late fall but smartphones using Merrifield aren't going to be out until early 2014, giving competing companies like NVIDIA, Qualcomm and Samsung more than enough time to respond. Intel did supply us with "forward looking" performance comparison that attempted to guess at the competitions next generation products but we weren't allowed to present that today.
I also don't know what to expect in the world of graphics - even though Silvermont and its completely re-architected design looks like a revolutionary step for mobile computing nothing was discussed on what graphics IPs might find their way into these products. Intel currently uses PowerVR designs for Medfield and Clovertrail but rumors abound that Bay Trail, the chip for tablets, will use Intel HD Graphics GPU designs. How much would be changed on that design will tell us a lot more about the potential success of these parts.
Soon we'll know more about the products that AMD has developed for the tablet market called Kabini. It would be impressive if AMD is able to come out of nowhere and build a competent, and readily available, mobile processor. Not only that, but we have solid expectations of what Kabini will offer consumers on the graphics side.
Is it possible that Silvermont is to Atom what Merom and Conroe were to the Pentium 4 desktop line, the beginning of a complete shift for Intel's development teams? Intel would like us to believe that is the case and that Silvermont and Airmont will do for them in the mobile space what the move to the Core architectures have done for Intel in the notebook and desktop markets. We'll finally be able to see this fall.