Semi-custom CPU
Qualcomm is ready in 2017 for its new flagship Snapdragon SoC, the 835.
With the near comes a new push for performance, efficiency and feature leadership from Qualcomm and its Snapdragon line of mobile SoCs. The Snapdragon 835 was officially announced in November of last year when the partnership with Samsung on 10nm process technology was announced, but we now have the freedom to share more of the details on this new part and how it changes Qualcomm’s position in the ultra-device market. Though devices with the new 835 part won’t be on the market for several more months, with announcements likely coming at CES this year.
Qualcomm frames the story around the Snapdragon 835 processor with what they call the “five pillars” – five different aspects of mobile processor design that they have addressed with updates and technologies. Qualcomm lists them as battery life (efficiency), immersion (performance), connectivity, and security.
Starting where they start, on battery life and efficiency, the SD 835 has a unique focus that might surprise many. Rather than talking up the improvements in performance of the new processor cores, or the power of the new Adreno GPU, Qualcomm is firmly planted on looking at Snapdragon through the lens of battery life. Snapdragon 835 uses half of the power of Snapdragon 801.
The company touts usage claims of 1+ day of talk time, 5+ days of music playback, 11 hours of 4K video playback, 3 hours of 4K video capture and 2+ hours of sustained VR gaming. These sound impressive, but as we must always do in this market, you must wait for consumer devices from Qualcomm partners to really measure how well this platform will do. Going through a typical power user comparison of a device built on the Snapdragon 835 to one use the 820, Qualcomm thinks it could result in 2 or more hours of additional battery life at the end of the day.
We have already discussed the new Quick Charge 4 technology, that can offer 5 hours of use with just 5 minutes of charge time.
Since we already knew that the Snapdragon 835 was going to be built on the 10nm process from Samsung, the first such high performance part to do so, I was surprised to learn that Qualcomm doesn’t attribute much of the power efficiency improvements to the move from 14nm to 10nm. It makes sense – most in the industry see this transition as modest in comparison to what we’ll see at 7nm. Unlike the move from 28nm to 14/16nm for discrete GPUs, where the process technology was a huge reason for the dramatic power drop we saw, the Snapdragon 835 changes come from a combination of advancements in the power management system and offloading of work from the primary CPU cores to other processors like the GPU and DSP. The more a workload takes advantage of heterogeneous computing systems, the more it benefits from Qualcomm technology as opposed to process technology.
Regardless of those traits, moving to a smaller process technology gives the 835 a smaller package size allowing for thinner phones, larger batteries and overall better feature sets.
Before diving into the CPU details, Qualcomm was adamant that the importance of their Symphony System Manager be pointed out. This very customizable – and specific to Snapdragon – technology is what Qualcomm uses to distribute work between the CPU cores, GPU, and DSPs. This isn’t how enthusiasts typically think about processors, but in the form factors that Qualcomm targets with the Snapdragon line it’s the best decision you can make. CPUs are great a serial tasks and sequential control, and the Hexagon DSP line is great for image and audio processing, object detection – anything that needs to be low power and real time. For its part the GPU is best for highly parallel tasks that have streaming data sets – graphics and composition.
The new Hexagon 682 DSP includes some new features with TensorFlow machine intelligence support and high efficiency vector extensions (HVX). Updated use-cases for Hexagon include visual inertial odometry (VIO), which combines data from a single camera and accelerometers in order to create accurate and real time poses for tetherless VR headset use, and the All-Ways Aware hub that keeps an eye on all the real-time data a phone has access to.
But let’s get to the new CPU, the Kryo 280. The Kryo 280 is the first processor built on ARM’s new semi-custom program called “Built on ARM Cortex Technology” that allows a partner like Qualcomm to take an off-the-shelf core (such as the Cortex-A73) and make modifications to it and rebrand it. This is a shift from the previous options of ARM cores or fully custom. Qualcomm, along with Apple and Samsung, had been the best examples of custom core designs for ARM SoCs, proving that you could do better with the added work of building your own CPU cores with an existing microarchitecture.
The result is an 8-core processor with four large cores and four smaller cores, similar to what we know as the ARM big.LITTLE design. The performance cores run up to 2.45 GHz, share 2MB of L2 cache and are 20% faster than the previous generation in a range of uses cases including app loads and VR. The smaller, efficiency cores will clock up to 1.9 GHz and have 1MB of L2 cache. While all eight cores have importance in the SoC, the chip spends 80% of the time running on the efficiency cores so the changes here can be more meaningful on total efficiency. Qualcomm did state that these cores would not be able to work at the same time, only the performance or efficiency cluster can be operating at one time.
When asked about the move to semi-custom, Qualcomm was a little defensive. The company needs to ensure that the audience doesn’t assume that because these are the “fully custom” parts that we have seen from Qualcomm over the years, that it means the Snapdragon 835 processor is compromised. Qualcomm said that it doesn’t “do things for political reasons” and that it looks at the technology landscape each iteration uses the best technology that lets them meet their goals each time. In the future, Qualcomm will again “focus on what is the most advantageous” indicating they could, COULD, return to fully custom cores with upcoming processor core generations.
Just what is Qualcomm customizing on the cores and what cores are they based on? Typically, Qualcomm isn’t going into much detail, not even telling us what cores are being modified. (I mentioned the Cortex-A73 above, which would typically be paired with the Cortex-A53 in a big.LITTLE configuration.) I was able to get some nuggets of information though. The efficiency core block has minimized transaction power with an increased L2 cache size. The bus interface on the cores was adjusted to fit into Qualcomm’s heterogeneous computing offloading scheme. Branch predictors were modified as well to better match past Kryo cores.
The memory controller on Snapdragon 835 is dual channel LPDDR4x capable of speeds up to 1866 MHz.
“Qualcomm, along with Apple
“Qualcomm, along with Apple and Samsung, had been the best examples of custom core designs for ARM SoCs”!
No Apple has a true custom core that is a twice as wide order superscalar design than Qualcomm’s semi-custom designs or ARM Holdings’ own reference core designs. The Apple Cyclone cores where an in-house design and no one has enough information on the Qualcomm 820 to verify its design as fully in house custom. Apple even beat ARM Holdings itself, the creator of the ARMv8A ISA, in getting a 64 bit custom ARMv8A ISA running micro-architecture to market.
Let’s stop with the just benchmarks and start making these “Custom” ARM based companies to offer the same necessary CPU core data sheets and technical information that the x86 makers do for their respective CPU cores. This lack of properly published custom ARM core specification listings is getting a bit too old now. Look at the core information provided by AMD for Ryzen/older, or Intel for SandyBridge/newer and older and even IBM for Power8/Power9 and ask yourselves what are the custom ARM core makers hiding, are these “custom cores” really custom or are they using mostly ARM holdings’ reference designs and misleading the markets.
Apple’s Cyclone is the last fully custom ARMv8A running Apple CPU core to be detailed and Anand Lal Shimpi was the last reporter outside of a pay wall with the skill set to write a proper custom ARM CPU core review. The Apple Cyclone was in no part by Apple properly explained to the market, it took Anand Lal Shimpi to do that.
Nvidia’s Really Custom Denver cores where properly presented at the Hot Chips Symposium so Nvidia is more like AMD, Intel, and IBM in that regard to not insulting the technically inclined with so much marketing driven drivel in the place of proper white papers that describe in great detail a REAL custom CPU core’s inner workings.
Benchmarks can be gamed, Show me the hardware!!!
There are no Reviews of many, to this day, of the custom ARM cores there are only Previews that remain only that the preview level for lack of proper reporting on the custom ARM CPU core designs!
You can go read Anand Lal Shimpi’s Apple Cyclone reviews and those where real reviews, not this eternal Preview stuff available on tech sites since Anand departed AnandTech!
Well, Anand ended up with a
Well, Anand ended up with a nice, mysterious and cushy job at Apple too.
Yes and Apple hired him to
Yes and Apple hired him to shut him up! So Anand had his price! I’m very surprised that Anand did not go on to found his own pay-walled publication but we do have the Linley Group’s Microprocessor Report that has been doing what Anand was doing since Anand himself was in diapers.
I do mean more so the Microprocessor Report when it was an independent publication, but it’s still fairly good under the Linley Group’s control, that is when Linley keeps their hands off of the reporting end of what the Microprocessor Report does best!
And Now “AnandTech” is owned by the same media outfit that owns Tom’s hardware! But AnandTech can still do a good article from time to time before the really good reporters are hired up away to the pay-walled professional Trade Journals like the Microprocessor Report and others. There was some particularly good reporting on the ARM Holdings A73 reference design cores and also on the Arm Holdings new Mali/Bifrost GPU micro-architecture and even Tom’s hardware can produce some good articles from time to time also but that is becoming rarer and rarer as time passes.
Currently under a very special Apple NDA, Anand Lal Shimpi is sealed away in carbonite inside a deep underground secret chamber below the under construction new Apple Space Ship Headquarters.
Well it looks like I’ll eat
Well it looks like I’ll eat some crow on the 820/custom bit but Charlie over at S/A is doing a rare for S/A deeper dive into the 835 in a rather long article that he occasionally does. So it’s a good read for more information not hidden behind S/A’s usual pay-wall!
“Qualcomm opens up a bit more on the 10nm Snapdragon 835 SoC”
http://semiaccurate.com/2017/01/03/qualcomm-opens-bit-10nm-snapdragon-835-soc/
Shilling works better when
Shilling works better when you’re not so obvious about it.