CPU Performance

While Denver’s architecture is something fascinating to study, it’s important to see how well this translates to the real world. Denver on paper is a beast, but in the real world there are a number of factors to consider, not the least of which is the effectiveness of NVIDIA’s DCO. We’ve laid out that Denver’s best and worst case scenarios heavily ride on the DCO, and for NVIDIA to achieve their best-case performance they need to be able to generate and feed Denver with lots and lots of well optimized code. If Denver spends too much time working directly off of ARM code or can’t do a good job optimizing the recurring code it finds then Denver will struggle. Meanwhile other important factors are in play as well, including the benefits and drawbacks of Denver’s two cores versus competing SoC’s quad A15/A57 configurations, and in thermally constrained scenarios Denver’s ability to deliver good performance while keeping its power consumption in check.

In order to test this and general system performance, we turn our suite of benchmarks that include browser performance tests, general system tests, and game-type benchmarks. As Denver relies on code-morphing to enable out of order execution and speculative execution, most of these benchmarks should be able to show ideal performance as loop performance in Denver is basically second to none. While most of these benchmarks are showing their age, they should be usable for valid comparisons until we move to our new test suite.

SunSpider 1.0.2 Benchmark  (Chrome/Safari/IE)

Kraken 1.1 (Chrome/Safari/IE)

Google Octane v2  (Chrome/Safari/IE)

WebXPRT (Chrome/Safari/IE)

Basemark OS II 2.0 - Overall

Basemark OS II 2.0 - System

Basemark OS II 2.0 - Memory

The Basemark System test seems to contribute quite strongly to how the Nexus 9 performs in the overall subtest. Given that this is a storage performance benchmark, it's likely that Basemark OS II has issues similar to Androbench on 5.0 Lollipop or that random I/O is heavily prioritized in this test.

Basemark OS II 2.0 - Graphics

There's a noticeable performance uplift in the graphics test, and although not exactly part of the CPU this does seem at least somewhat plausible as GPU driver updates can improve performance over time.

Basemark OS II 2.0 - Web

Overall, performance seems to be quite checkered, although improved from our initial evaluation of the Nexus 9. Unfortunately, even in benchmarks where the DCO should be able to easily unroll loops to achieve massive amounts of performance, we see inconsistent performance in Denver. This may come down to an issue with the DCO, or even more simply the fact that Denver is spending more time than it would like to directly executing ARM code as opposed to going through the DCO.

In this case looking at the SunSpider and Kraken javascript benchmarks offers an interesting proxy case for exactly that scenario. SunSpider on modern CPUs executes extremely quickly, so quickly that the individual tests are often over in only a couple of dozen of milliseconds. This is a particularly rough scenario for Denver, as it doesn’t provide Denver with much time to optimize, even if the code is run multiple times. Meanwhile Kraken pushes many similar buttons, but its tests are longer, and that gives Denver more time to optimize. Consequently we find that Denver’s SunSpider performance is quite poor – underperforming even the A15-based Tegra K1-32 – while Denver passes even the iPad Air 2 in Kraken.

Ultimately this kind of inconsistent performance is a risk and a challenge for Denver. While no single SoC tops every last CPU benchmark, we also don’t typically see the kind of large variations that are occurring with Denver. If Denver’s lows are too low, then it definitely impacts the suitability of the SoC for high-end devices, as users have come to expect peppy performance at all times.

In practice, I didn't really notice any issues with the Nexus 9's performance, although there were odd moments during intense multitasking where I experienced extended pauses/freezes that were likely due to the DCO getting stuck somewhere in execution, seeing as how the DCO can often have unexpected bugs such as repeated FP64 multiplication causing crashes. In general, I noticed that the device tended to also get hot even on relatively simple tasks, which doesn't bode well for battery life. This is localized to the top of the tablet, which should help with user comfort although this comes at the cost of worse sustained performance.

SPECing Denver's Performance GPU and NAND Performance
Comments Locked

169 Comments

View All Comments

  • Mondozai - Wednesday, February 4, 2015 - link

    No offence but how relevant is this review so many months after release?
    You guys dropped the ball on this one. We're also still waiting for the GTX 960 review.

    What has happened to Anandtech...
  • LocutusEstBorg - Wednesday, February 4, 2015 - link

    There's no Anand.
  • nathanddrews - Wednesday, February 4, 2015 - link

    That's the only change I've noticed.
  • Morawka - Wednesday, February 4, 2015 - link

    and no Brian Klug
  • nathanddrews - Wednesday, February 4, 2015 - link

    Yeah, but that was earlier.
  • Ryan Smith - Wednesday, February 4, 2015 - link

    "What has happened to Anandtech..."

    Nothing has happened to AnandTech. We're still here and working away at new articles.=)

    However this article fell victim to bad timing. The short story is that I was out sick for almost 2 weeks in December, which meant this got backed up into the mess that is the holidays and CES.

    As for how relevant it is, it is still Google's premiere large format tablet and the only shipping Denver device, both of which make it a very interesting product.
  • Jon Tseng - Wednesday, February 4, 2015 - link

    It's fine to be late (although maybe not as late as the Razer Blade 2014 review!). Better to have late, differentiated content than early, commoditised content. Whether the review like's the colour of a tablet's trim is of limited interest for me; the details of Denver code-morphing are.

    Actually my worry is that under new ownership Anandtech might be pushed to go down the publish early/get click views route vs. the publish late/actually deliver something useful. Hopefully it won't come to this, but this is what historically happens... :-(
  • Operandi - Thursday, February 5, 2015 - link

    Being there on day one is not a huge deal but its certainly not ok be as late as this review is or the still MIA 960 review. If you are going to be late you better be brining something new to the table to justify not being there in the same time frame as your peers. This is so laughably late its almost embarrassing to release it at all at this point.

    Tech journalism like most other markets is competitive and there are lots of other very competent publications out there competing for the same readers. Personally I've already gotten all the Nexus 9 information elsewhere so this review is of no value to me whatsoever. The same goes for the 960 review when/if that review ever shows up.
  • akdj - Wednesday, February 11, 2015 - link

    Not sure where you've seen such an extensive write up and dissection of Denver, but I certainky haven't. Nor were the N9/6 widely available until the holidays were over. Like a month ago
    For every 10,000,000 iPads produced, HTC is probably knocking out 10,000
    Excellent review, write up and information about the 'other 64bit' option.
  • Taneli - Wednesday, February 4, 2015 - link

    Timing is secondary for a deeply technical article like this here. You guys did exactly the right thing, reporting when the device was announced and waited for the review to be done before publishing. Also, having people out sick in a small team is something you really can't do that much about. I hope you're well now.

    The article itself was superb. Thanks for the read and keep up the good work.

Log in

Don't have an account? Sign up now