Yesterday we presented the first results of Qualcomm's Krait based MSM8960 SoC. While we still await the first Krait based phones (widely expected to begin shipping sometime in Q2), courtesy of Qualcomm's MSM8960 Mobile Development Platform we were able to get a good idea of the upper bound for Krait and MSM8960 performance. I mention it's the upper bound because, at least in the past, MDP performance hasn't corresponded directly to shipping device performance. There was a pretty big delta between MSM8660 MDP performance and phones that used the MSM8660. Qualcomm tells us that this time around things are going to be different. Qualcomm is expecting a much narrower (nonexistent?) gap between the MSM8960 development platform and phones that use MSM8960 silicon. One major difference between the MSM8960 MDP and our earlier MSM8660 MDP was the state of the CPU governor. In the earlier MDP the governer was set to max performance, always delivering the CPU's maximum clock frequency. With the MSM8960 platform the governor was set to ondemand, allowing for variable CPU speeds depending on what the OS requests of the device. The ondemand setting is in-line with what we can expect device manufacturers to use when they ship phones. All of this goes to say that while we have a good handle of what Krait and the MSM8960 are capable of, there are still a lot of unknowns.

While it's true that shipping performance remains to be seen, some of the deltas we saw between MSM8960 and the current competition were so great that even a much slower implementation in a shipping phone would still be significantly faster than anything else out today.

We left our MSM8960 investigation with two major unknowns. The first was power consumption. We still haven't been able to get Qualcomm's Trepn tool running on the MSM8660 MDP, which has always been a bit finicky. To get a true feel for MSM8960 battery life we will have to wait for shipping devices. The other major unknown was really how MSM8960 stacks up against NVIDIA's Tegra 3.

Tegra 3 was everything Tegra 2 should have been. We got higher clocks, NEON support and a much faster GPU. The only thing missing from Tegra 3 was a dual channel memory interface. We were happy with Tegra 3 on ASUS' Eee Pad Transformer Prime, but in less than a week we'll get to meet some of the first smartphones based on T3 silicon.

Armed with the Eee Pad Transformer Prime (updated to Ice Cream Sandwich) we're able to get a rough idea of how these two heavyweights will compare. The same caveats that applied to the MDP apply to our Tegra 3 platform as well. Since we are using a tablet we're obviously dealing with a higher TDP than what you'll find in a phone. The comparison today is largely academic and naturally shipping devices may be better or worse that these two representatives. With the disclaimers out of the way, let's get to the comparison.

CPU Performance: Preferring Single vs. Multithreaded Performance

The MSM8960 features two Krait cores compared to the four ARM Cortex A9 cores in NVIDIA's Tegra 3. While the A9 is a very power efficient core, Krait offers a much wider front end, wider execution back end, faster FPU and an improved cache/memory interface. All of these factors together combined with similar clock speeds to what Tegra 3 is able to hit should result in better absolute performance in single or lightly threaded applications. As video decode and transcode are both fully offloaded in all modern SoCs, finding workloads that scale well across more than two cores is difficult. We noted this in our Eee Pad Transformer Prime review - it's just not easy coming up with current apps that scale well to four ARM cores. That's not to say that there are no advantages to more than two cores, but you're more likely to get a benefit from two faster cores vs. four slower ones.



NVIDIA's saving grace is the fact that it did ramp up A9 clock speed very high in Tegra 3, and it has that handy companion core 4-PLUS-1 architecture to keep power consumption low throughout very light workloads. There's also the fact that while very few smartphone apps will peg four cores constantly, there are periods of time when you'll see more than two cores in use. Multitasking, although more likely to happen in significant amounts on a tablet, can also increase usage of the third and fourth cores on Tegra 3.

We'll start with Linpack, our heaviest floating point/cache/memory bandwidth test:

Linpack - Single-threaded

Single threaded floating point performance is obviously a strength of the MSM8960 and Krait. Qualcomm tells us that Krait is able to multi-issue floating point instructions, something that the Cortex A9 cannot do. The MSM8960 memory controller also appears to be more efficient than previous designs, contributing to the magnitude of the win here.

Move to more threads and the situation doesn't change dramatically, although Tegra 3 is obviously far more competitive thanks to its sheer core count:

Linpack - Multi-threaded

Javascript performance can be multithreaded at times but most of the benchmarks we run don't scale incredibly well beyond two cores. Making matters worse is the fact that SunSpider performance regressed on the Eee Pad Transformer with the latest update to ICS. I've included the old Honeycomb results as a reference for where things should be. Keep in mind that the Honeycomb browser on the Eee Pad Transformer was very heavily optimized for Tegra 3. It's possible that the same degree of optimizations just aren't present in the ICS version yet.

SunSpider Javascript Benchmark 0.9.1 - Stock Browser

Browsermark tells a different story. Here the Tegra 3 based Transformer Prime is actually able to be slightly faster than the MSM8960. The margin of victory is small enough to be a wash, but the fact that NVIDIA is able to remain competitive is important.


Basemark OS echoes more of what we'd expect. In the overall score the MSM8960 is around 50% faster than the Tegra 3 based tablet. Even if the MSM8960 MDP is unrealistically fast for a Krait platform, it's likely that we'll still see a Krait advantage.

Basemark OS - System
  HTC Rezound Galaxy Nexus ASUS Transformer Prime MDP MSM8960
System Overall Score 658 538 602 907
Simple Java 1 298 loops/s 210 loops/s 240 loops/s 375 loops/s
Simple Java 2 7.28 loops/s 8.61 loops/s 7.27 loops/s 10.8 loops/s
SMP Test 35.3 loops/s 49.2 loops/s 81.2 loops/s 64.4 loops/s
100K File (eMMC->SD) 6.49 mB/s 9.52 mB/s 11.0 mB/s 8.64 mB/s
100K File (SD->eMMC) 33.0 mB/s 17.8 mB/s 14.5 mB/s 39.8 mB/s
100K File (eMMC->eMMC) 37.8 mB/s 34.5 mB/s 29.7 mB/s 48.9 mB/s
100K File (SD->SD) 8.47 mB/s 8.30 mB/s 8.06 mB/s 12.7 mB/s
Database Operation 10.0 ops/s 5.73 ops/s 4.56 ops/s 19.4 ops/s
Zip Compression 0.509 s 0.848 s 0.637 s 0.561 s
Zip Decompression 0.097 s 0.206 s 0.089 s 0.073 s

Most of the Basemark tests are lightly threaded, but looking at the SMP test gives you another example of Tegra 3's strengths given the right workload. With the right application, Tegra 3 can be faster than the MSM8960, however it's still our opinion that you're more likely to find a lightly threaded workload on a smartphone than you are going to encounter something that scales well to four cores.

GPU Performance
Comments Locked


View All Comments

  • Alaa - Thursday, February 23, 2012 - link

    How about running more than one of these processes at once? Wouldn't 4 cores get more benefit?
  • MonkeyPaw - Thursday, February 23, 2012 - link

    The thing is, Tegra 3 is still on a single channel of memory. This was one of the design decisions that made me really question how valuable it was going to 4+1 cores. I think this is probably a very big reason it can't best Krait's 2 cores in any multi-threaded tests. I think they would have been better off going to a 2 core dual channel setup with improved graphics, and saved the move to quad core until Tegra 4.
  • tipoo - Thursday, February 23, 2012 - link

    The single channel memory was my first thought also. Four cores active at once plus a GPU all sharing one channel sounds like a terrible idea, that's probably a significant factor.
  • piroroadkill - Thursday, February 23, 2012 - link

    Your two cores crapped all over Tegra 3's 5.
    My god, Tegra 3 looks bloated compared to MSM8960.
    I eagerly await devices of all types with this chipset! Bravo!
  • rahvin - Friday, February 24, 2012 - link

    Just like Tegra and Tegra2 nVidia made design decisions to push paper specs over real performance and power savings. Personally other than a few token devices I expect Tegra3 to be a no show once Krait and Omap5 hit the scene.
  • LetsGo - Thursday, February 23, 2012 - link

    Looking at the number in the linpack benchmark I would expect the results to scale less than linearly, MDP MSM8960 = 107 MFLOPS single threaded, 218MFLOPS Multi- threaded!

    That chip has the smarts Impressive!
  • metafor - Thursday, February 23, 2012 - link

    Since two cores share an L2 cache, it's very possible that one core's prefetch also feeds another. I don't think the Linpack program is very intelligent in this regard :/
  • tipoo - Thursday, February 23, 2012 - link

    The SGX 543MP2 in the iPad 2 pushes 148 in GLbench Pro offscreen. A year after its first début, why is no one else using it yet? I know the A5 die size is much larger than comparable SoCs because of it, but this here shows everyone else is focusing on CPU performance over GPU, and I think after a certain point the GPU will matter more for rich content.

    And the iPad 3 will be launching within months no doubt, some saying with a 543MP4 like the Vita has, which would make a large lead into an enormous one. Seriously, what's going on with everyone else?
  • Death666Angel - Thursday, February 23, 2012 - link

    iOS unfortunately is pretty much alone in the mobile gamer department. Android is not really pushing mobile games as much. So big beefy GPUs are not that marketable. :-(
  • Dribble - Thursday, February 23, 2012 - link

    I always thought of tegra 3 as a stepping stone between last years standard of dual A9's @ 40nm to tomorrows standard of quad A15's @ 28nm.

    It looks like we have most of tomorrows chip today as we have a dual A15 equivalent @ 28nm. However is it really shipping soon - 28nm production isn't exactly where it needs to be yet for cheap mobile phone chips? Are the final released products going to be that far ahead of the 28nm A15's?

Log in

Don't have an account? Sign up now