Earlier tonight LG announced its Optimus G, the company's first Qualcomm Fusion 3 based smartphone (APQ8064 + MDM9615). If you don't follow Qualcomm's naming system closely, we're talking about a quad-core 28nm Krait SoC with Adreno 320 graphics core paired with its very popular 28nm LTE baseband. The CPU cores can independently clock up to 1.5GHz and users can even manually shut off a pair of cores to save power. 

Brian got some hands on time with the Optimus G earlier tonight, but that just extended into a benchmarking session. While Brian spends more time with the Optimus G, he did send over a bunch of preliminary performance results from the device. We've seen APQ8064 in Qualcomm's MDP/T reference tablet, but this is the first time we've been able to gauge performance in a (nearly) shipping device. 

I'm still not pleased with the current state of Android benchmarking but we'll have to make do with what we've got.

First off is Linpack, a reasonable indicator of heavy FP performance and memory bandwidth, but an even better tool for showing single and multi-core scaling.

Linpack - Single-threaded

Linpack - Multi-threaded

Qualcomm's Krait architecture has always done very well in this test, but the move to four cores doesn't actually show perfect scaling here. It's unclear if the test is simply running too quickly to produce a consistently high result, or if we're just running into issues feeding all four cores running at 1.5GHz because the 1 to 4 thread scaling is only around 2.45x here (vs 4x ideal). Either way we do see scaling beyond what you get from two cores with the quad-core APQ8064.

Next up we have our two standard browser benchmarks: Sunspider and Browsermark. These two tests use LG's custom browser and measure aspects of JS performance:

SunSpider Javascript Benchmark 0.9.1 - Stock Browser

Sunspider is lightly threaded and thus doesn't see huge scaling going to four cores. In fact, in this case we're not seeing any real improvement over the dual-core Krait based devices from HTC. It's unclear how much of the Optimus G's performance is due to LG's browser/software stack vs. the underlying hardware.

BrowserMark

BrowserMark doesn't look great and the Optimus G's performance is almost certainly due to LG's own browser code. Qualcomm's reference software stack can provide great performance, but it's up to the individual OEM to take advantage of it. 

Finally we have our two GLBenchmark 2.5 GPU tests: Egypt Classic and HD, both run in their offscreen 1080p modes without Vsync enabled. Here the Adreno 320 is really able to shine:

GLBenchmark 2.5 - Egypt HD (Offscreen 1080p)

GLBenchmark 2.5 - Egypt Classic (Offscreen 1080p)

Performance looks awesome in both tests. The Egypt HD numbers are bordering on reasonably smooth and the classic numbers are great of course. Both of these tests run at resolutions higher than the Optimus G's native 1280 x 768 panel. At native res you're looking at 37.5 fps for Egypt HD and 59.8 fps for Egypt Classic, both extremely smooth.

If anything it's the GPU performance that's most exciting about the Optimus G. Most Android applications don't show great scaling from two to four cores, but the boost you get from the Adreno 320 will definitely be apparent in visually demanding games.

POST A COMMENT

20 Comments

View All Comments

  • Phelerox - Tuesday, September 18, 2012 - link

    Any theories as to why quad core Krait performs almost twice the MFLOPS in single-threaded Linpack that the dual-core Krait pushes? Reply
  • AndreiLux - Tuesday, September 18, 2012 - link

    Anand is still using the Dalvik based Linpack test which is absolutely useless in comparing between different phones as the OS and whatever platform code can make huge differences. I told Brian about StabilityTest's native NDK implementation of Linpack several months ago but they still release reviews with the flawed Linpack. Reply
  • toyotabedzrock - Tuesday, September 18, 2012 - link

    Maybe they fixed the memory bandwidth problem and licensed the IP to Apple.
    If they can they should run geekbench on it to see how it compares to the previous chips.
    Reply
  • dishayu - Tuesday, September 18, 2012 - link

    I have the same question as Phelerox. It oddly seems that the "single threaded" mode simply uses half the cores available. In most cases, the performance just scales up to double when "multi threaded" mode is used. This holds true for Exynos quad and Tegra 3 as well.

    About the International One S... Is it the krait based international model or the "limited" 1.7 Ghz Scorpion based One S? I ask because it wouldn't make much sense to have an international version and a T mobile version in benchmarks if they were using exactly the same hardware.
    Reply
  • OCedHrt - Tuesday, September 18, 2012 - link

    Regarding your second question, the international One S is not limited to the 1.7 dual core S3. Only certain international markets are stuck with the 1.7 S3. Reply
  • dishayu - Tuesday, September 18, 2012 - link

    Yup. When i said "limited" 1.7 scorpion, i meant limited to certain markets. Reply
  • dishayu - Tuesday, September 18, 2012 - link

    Kindly ignore the first part of my statement. I has just woken up and had a brainfart while reading those charts. The performance does scale up linearly from one core to 4 making Phelerox's question even more interesting. Reply
  • cmikeh2 - Tuesday, September 18, 2012 - link

    I don't know why single-threaded performance is so high but my guess is the SoC is memory bandwidth-limited once it reaches full utilization of all four cores. It's possible that Qualcomm increased the memory cache sizes in the S4 Pro SoC's and that's why we're seeing the big performance boost in the single-threaded run as well. Just my two cents. Reply
  • cmikeh2 - Tuesday, September 18, 2012 - link

    After looking for details on the differences between the 8960 and the 8064 I've decided that that's probably wrong and should be disregarded as such. Now I'm stumped too. Reply
  • cmikeh2 - Tuesday, September 18, 2012 - link

    Actually the L2 cache for the APQ8064 is 2 MB compared to the 1 MB for the MSM8960 but I don't know what the effect of that would be. Reply

Log in

Don't have an account? Sign up now