Qualcomm Adreno 4xx

If they're committing to the expense of two separate pieces of silicon to target those different tiers of the market, they might as well tailor each to carry relatively large differences.

 

Qualcomm seems to be aiming 2GHz tops for the 808/810 on the A57 cores: http://www.qualcomm.com/snapdragon/processors/810

That's quite conservative and less than I expected.

 

Let's hope A57s are efficient enough to not need higher frequencies.

 

Anandtech's Snapdragon 805 performance preview (using Qualcomm's reference 10.6" MDP tablet) is up: http://www.anandtech.com/show/8035/qualcomm-snapdragon-805-performance-preview

The GPU has a nice ~ 20-50% performance boost compared to S800, but the CPU performance boost is negligible. To be quite honest, Cortex A15 looks to be significantly ahead of Krait 450 with both fabricated on 28nm HPM. And with 64-bit Cortex A57 being even higher performance than Cortex A15, with availability next year, it's not too surprising that Qualcomm will be going that route for their high end SoC in 2015.

 

Time to bless my laurels then:

http://forum.beyond3d.com/showpost.php?p=1847662&postcount=2348

Ok Anand got 17.7 fps

 

Considering it has nearly twice the bandwidth (this chip screams tablet to me btw) I'm not impressed (only talking about the gpu, that the cpu would be quite disappointing we already knew). Ok so it's got the next-gen feature set just like Tegra K1, and maybe with some driver improvements they could get competitive there but that they needed a 128bit interface to pull it off is a letdown imho.

 

It was quite some time ago since QCOM stated that the 420 will have roughly 40% higher arithmetic efficiency compared to the Adreno330 so no one really needed a crystal ball to see how its performance will look like roughly on the GPU level.

It's bandwidth benefit shows in Anand's fillrate and alpha blending results in Gfxbench3.0; what's disappointing (yet nothing new either) is the driver overhead score which is amongst the worst of its class.

In any case if that performance level can be reached in a smartphone (Anand's numbers are from a reference platform aren't they?), then roughly 18 fps in Manhattan isn't exactly a score to dismiss that easily, unless of course you're expecting the GK20A GPU in Tegra K1 to do significantly better in a smartphone.

 

That's true but I thought not all of these tests were completely ALU bound. Hence with such a big bandwidth increase "only" 40% faster doesn't sound so great to me.

Yes but I'd have thought it would make some difference in some "real world" workloads. Well it does I guess just in nothing which was benchmarked yet.

I'm not entirely sure how representative this number really is, but you could consider it a good thing as the chip might have potential to get faster with better drivers .

Yes if it can (mostly) keep the performance in a smartphone that would be quite ok. Doesn't have an integrated modem though so it seems doubtful it's going to be popular in smartphones. And the next chip with integrated modem (S808) only has a adreno 418 and goes back to 64bit memory interface, which makes me wonder if the gpu is going to be faster at all compared to a adreno 330.

 

Anand's numbers are from a reference MDP 10.6" tablet running on AC power. So it's impossible to precisely say what we can expect from S805 in a smartphone form factor. The S805 GPU is up to ~ 20% more power efficient than the S800 GPU according to Qualcomm, so the TK1 GPU is clearly more power efficient in comparison.

 

When a benchmark is ALU bound and the arithmetic increase is only 40% worth then there's no way around it.

Look at the alpha blending and fillrate results of the 420 against the 330; both should have 8 TMUs each and the first should be clocked slightly lower.

It doesn't do anything special any of the existing applications wouldn't do; it just measures performance with a lot of (a realistic amount according to Kishonti) state changes in order to see how big the CPU overhead is.

Usually at the start of each platform performance is somewhat lower in smartphones only to catch up through future silicon revisions; we'll see.

As long as there's neither a S805 nor a K1 smartphone compared by an independent source I wouldn't jump to conclusions yet; typically yes as I mention above at first performance is lower but catches up through chip revisions. The 330 didn't start out at today's 580MHz frequency either and the 420 frequency according to my own estimates should be in the 550MHz region. You should keep in mind that the Adreno420 has most likely the same amount of SIMD lanes as the GK20A GPU. The 420 could have in favour a somewhat higher frequency and against it it's rather mediocre drivers and end up more or less in the same neighborhood as a hypothetical K1 smartphone.