SoC Architecture: NVIDIA's Denver CPU

It admittedly does a bit of a disservice to the rest of the Nexus 9 both in terms of hardware and as a complete product, but there’s really no getting around the fact that the highlight of the tablet is its NVIDIA-developed SoC. Or to be more specific, the NVIDIA-developed Denver CPUs within the SoC, and the fact that the Nexus 9 is the first product to ship with a Denver CPU.

NVIDIA for their part is no stranger to the SoC game, now having shipped 5 generations of Tegra SoCs (with more on their way). Since the beginning NVIDIA has been developing their own GPUs and then integrating those into their Tegra SoCs, using 3rd party ARM cores and other 1st party and 3rd party designs to fully flesh out Tegra. However even though NVIDIA is already designing some of their own IP, there’s still a big leap to be made from using licensed ARM cores to using your own ARM cores, and with Denver NVIDIA has become just the second company to release their own ARMv8 design for consumer SoCs.

For long time readers Denver may feel like a long time coming, and that perception is not wrong. NVIDIA announced Denver almost 4 years ago, back at CES 2011, where at the time they made a broad announcement about developing their own 64bit ARM core for use in wide range of devices, ranging from mobile to servers. A lot has happened in the SoC space since 2011, and given NVIDIA’s current situation Denver likely won’t be quite as broad a product as they first pitched it as. But as an A15 replacement for the same tablet and high performance embedded markets that the TK1-32 has found a home in, the Denver-based TK1-64 should fit right in.

K1-64 Die Shot Mock-up (NVIDIA)

Denver comes at an interesting time for NVIDIA and for the ARM SoC industry as a whole. Apple’s unexpected launch of the ARMv8 capable Cyclone core in 2013 beat competing high-performance ARMv8 designs by nearly a year. And overall Apple set a very high bar for performance and power efficiency that is not easily matched and has greatly impacted the development and deployment schedules of other ARMv8 SoCs. At the same time because Cyclone and its derivatives are limited to iOS devices, the high-performance Android market is currently served by a mix of ARMv7 designs (A15, Krait, etc) and the just recently arrived A57 and Denver CPUs.

Showcasing the full scope of the ARM architecture license and how many different designs can execute the same instruction set, none of these ARMv8 CPUs are all that much alike. Thanks to its wide and conservatively clocked design, Apple’s Cyclone ends up looking a lot like what a recent Intel Core processor would look like if it were executing ARM instead of x86. Meanwhile ARM’s A57 design is (for lack of a better term) very ARMy, following ARM’s own power efficient design traditions and further iterating on ARM’s big.LITTLE philosophy to pair up high performance A57 and moderate performance A53 cores to allow a SoC to cover a wide power/performance curve. And finally we have Denver, perhaps the most interesting and certainly least conventional design, forgoing the established norms of Out of Order Execution (OoOE) in favor of a very wide in-order design backed by an ambitious binary translation and optimization scheme.

Counting Cores: Why Denver?

To understand Denver it’s best to start with the state of the ARM device market, and NVIDIA’s goals in designing their own CPU core. In the ARM SoC space, much has been made of core counts, both as a marketing vehicle and of value to overall performance. Much like the PC space a decade prior, when multi-core processors became viable they were of an almost immediate benefit. Even if individual applications couldn’t yet make use of multiple cores, having a second core meant that applications and OSes were no longer time-sharing a single core, which came with its own performance benefits. The OS could do its work in the background without interrupting applications as much, and greedy applications didn’t need to fight with the OS or other applications for basic resources.

However also like the PC space, the benefits of additional cores began to taper off with each additional core. One could still benefit from 4 cores over 2 cores, but unless software was capable of putting 3-4 cores to very good use, generally one would find that performance didn’t scale well with the cores. Compounding matters in the mobile ecosystem, the vast majority of devices run apps in a “monolithic” fashion with only one app active and interacting with the user at any given point in time. This meant that in absence of apps that could use 3-4 cores, there weren’t nearly as many situations in which multitasking could be employed to find work for the additional cores. The end result has been that it has been difficult for mobile devices to consistently saturate an SoC with more than a couple of cores.

Meanwhile the Cortex family of designs coming from ARM have generally allowed high core counts. Cortex-A7 is absolutely tiny, and even the more comparable Cortex-A15 isn’t all that big on the 28nm process. Quad core A15 designs quickly came along, setting the stage for the high core count situations we previously discussed.

This brings us to NVIDIA’s goals with Denver. In part due to the issues feeding 4 cores, NVIDIA has opted for a greater focus on single-threaded performance than the ARM Cortex designs they used previously. Believing that fewer, faster cores will deliver better real-world performance and better power consumption, NVIDIA set out to build a bigger, wider CPU that would do just that. The result of this project was what NVIDIA awkwardly calls their first “super core,” Denver.

Though NVIDIA wouldn’t know it at the time it was announced in 2011, Denver in 2015 is in good company that helps to prove that NVIDIA was right to focus on single-threaded performance over additional cores. Apple’s Cyclone designs have followed a very similar philosophy and the SoCs utilizing them remain the SoCs to beat, delivering chart-topping performance even with only 2 or 3 CPU cores. Deliver something similar in performance to Cyclone in the Android market and prove the performance and power benefits of 2 larger cores over 4 weaker cores, and NVIDIA would be well set in the high-end SoC marketplace.

Performance considerations aside, for NVIDIA there are additional benefits to rolling their own CPU core. First and foremost is that it reduces their royalty rate to ARM; ARM still gets a cut as part of their ISA license, but that cut is less than if you are also using ARM licensed cores. The catch of course is that NVIDIA needs to sell enough SoCs in the long run to pay for the substantial costs of developing a CPU, which means that along with the usual technical risks, there are some financial risks as well for developing your own CPU.

The second benefit to NVIDIA then is differentiation in a crowded SoC market. The SoC market has continued to shed players over the years, with players such as Texas Instruments and ST-Ericsson getting squeezed out of the market. With so many vendors using the same Cortex CPU designs, from a performance perspective their SoCs are similarly replaceable, making the risk of being the next TI all the greater. Developing your own CPU is not without risks as well – especially if it ends up underperforming the competition – but played right it means being able to offer a product with a unique feature that helps the SoC stand out from the crowd.

Finally, at the time NVIDIA announced Denver, NVIDIA also had plans to use Denver to break into the server space. With their Tesla HPC products traditionally paired x86 CPUs, NVIDIA could never have complete control over the platform, or the greater share of revenue that would entail. Denver in turn would allow NVIDIA to offer their own CPU, capturing that market and being able to play off of the synergy of providing both the CPU and GPU. Since then however the OpenPOWER consortium happened, opening up IBM’s POWER CPU lineup to companies such as NVIDIA and allowing them to add features such as NVLink to POWER CPUs. In light of that, while NVIDIA has never officially written off Denver’s server ambitions, it seems likely that POWER has supplanted Denver as NVIDIA’s server CPU of choice.

Introduction Designing Denver
Comments Locked


View All Comments

  • PC Perv - Wednesday, February 4, 2015 - link

    It is clear, even though you did not say, why no one other than NV and Google will use Denver in their products. Thank you for the coherent review, Ryan.

    P.S. I can't wait for the day SunSpider, Basemark, and WebXPRT disappear from your benchmark suit.
  • jjj - Wednesday, February 4, 2015 - link

    You always make those kind of claims about dual core vs more cores but you have never attempted to back them up with real world perf and power testing.
    In real use there are alerts and chats and maybe music playing and so on. While your hypothesis could be valid or partially valid you absolutely need to first verify it before heavily insisting on it and accepting it as true. Subjective conclusions are just not your style is it, you test things to get to objective results.
    And it wold be easy you already have "clean"numbers and you would just need to run the same benchmarks for perf and power with some simulated background activity to be able to compare the differences in gains/loses.
  • PC Perv - Wednesday, February 4, 2015 - link

    Where would you put the performance of "backup" ARM-only part of Denver? Cortex-A7? Is it measurable at all?

    Also, why don't Samsung use F2FS for their devices? I thought it was developed by them.
  • abufrejoval - Wednesday, February 4, 2015 - link

    While the principal designer seems to be a Korean, I'm not sure he works for Samsung, who typically used Yet Another Flash File System (YAFFS).
  • Ryan Smith - Wednesday, February 4, 2015 - link

    It's not measurable in a traditional sense, as the DCO will kick in at some point. However I'd say it's somewhere along the lines of A53, though overall a bit better.
  • Shadowmaster625 - Wednesday, February 4, 2015 - link

    The design philosophy of the DCO does make a lot of sense. When your mobile device starts to bog down and you start cursing at it, what is it usually doing? It is usually looping or iterating through something. The DCO wont help with small blocks of code that execute in 500uS, but you dont need help with that sort of code anyway. What you want to improve is exactly the type of code the DCO can improve: the kind of code that takes several dozen milliseconds (or more) to execute. That is when you begin to notice the lag in your cpu.
  • mpokwsths - Wednesday, February 4, 2015 - link

    Joshua & Ryan,

    please update the charts with the bench results of the newer version of Androbench 4:
    (I had previously commented on the fact that you can't safely compare the i/o results of different OS AND different bench apps).

    Androbench 4 is redesigned it to use multiple i/o threads (as a proper i/o bench app should have) and produces vastly improved results on both Lollipop and earlier Android devices.

    You will not be able to compare the newer results with older ones, but at least it will put an end to this ridiculus ι/ο performance difference between iOS and Android, the one you persistently -but falsly- keep projecting.
  • Andrei Frumusanu - Wednesday, February 4, 2015 - link

    I tested this out on several of my devices and could see only minor improvements, all within 10%. The performance difference to iOS devices does not seem to be a dupe at all.
  • mpokwsths - Wednesday, February 4, 2015 - link

    My results strongly disagree with you:
    Nexus 5: Seq Write: 19MB/s --> 55 MB/s
    Rand Write: 0.9 --> 2.9 MB/s

    Sony Z3 Tablet: Seq Write: 21 MB/s --> 53 MB/s
    Rand Write: 1,6 MB/s --> 8MB/s
    Seq Read: 135 MB/s --> 200MB/s

    I can upload pics showing my findings.
  • mpokwsths - Wednesday, February 4, 2015 - link

    Meet the fastest Nexus 5 in the world:


Log in

Don't have an account? Sign up now