Qualcomm Snapdragon 800 (MSM8974) Performance Preview

[Exophase]

I agree that these level of performance are probably not achieved in < 3W on that phone, but just because the phone and tablet perform the same doesn't mean they're both using > 5W. I reiterate: there's a physical limit to what that 4.3" MDP phone is capable of dissipating safely. They wouldn't have to check device temperatures to know the thing got too hot to hold. How much can this phone really handle? 5W for the SoC has got to be pushing the upper limits (considering at least 1W for everything else - display, memory, etc). Honestly the only reason I'm saying this now and not earlier is because I didn't even know the MDP phone was reviewed and got the same results as the tablet; had I known that I would have been a lot more skeptical that the GPU was using 6W over 2 minutes.

Has there even been a Qualcomm tablet MDP? They haven't had much presence in tablets to begin with. That's probably going to start changing a lot.

3W wouldn't be doubling the efficiency over 225 at the very least. I said it in an earlier post, scale the performance numbers given vs the power numbers and it'd be 4.7W. That'd be increasing efficiency by 1.57x. Note that Adreno 225 was just an overclocked Adreno 220 which is a pretty old uarch. Now, I'm not actually saying 3W for the GPU - I could see it taking around 4W and the whole thing still fitting in a 5W envelope. The CPU cores wouldn't necessarily need to be clocked that aggressively for GLBench 2.5. This isn't that different from scenarios we've seen with Exynos 5250 where the GPU was running full tilt and the CPU cores at 800MHz.

 

[Gideon]

while staying on the same process...

While I agree overall, that these gains are unlikely without similar gains in TDP, i just wanted to point out that it isn't exactly the same process:

While I'm aware, that this is a marketing slide, the power draw still seems a bit lower on HPm vs LP, especially at higher frequencies.

 

[Khato]

Not certain how much the phone can handle or how hot it got since none of the coverage that I read mentioned anything about temperatures. Which is something that piqued my curiosity as normally there would have at least been some comments about the MDP's being cold/warm/hot after running the benchmarks. (And yes, that's somewhat a poke to be proven wrong as I'll make no claim of reading all the coverage.) Regardless, the same performance in both phone and tablet MDP when the Fudzilla article claims there should be a 3W vs 5W TDP difference between them calls into question whether those limits were in place.

The APQ8064 was also demonstrated in a (smaller) tablet MDP - http://www.anandtech.com/show/6112/...agon-s4-apq8064adreno-320-performance-preview

As for scaling the performance numbers versus Adreno 225, how did you arrive at Adreno 330 only being 4.7x higher performance? The highest GLBenchmark 2.7 score I see for Adreno 225 is 4.6 fps, and the Dell XPS 10 that we obtained the 1W figure for Adreno 225 only gets 3.2 fps. (For which I'd blame it running Windows RT instead of Android.) Even in the best case, Adreno 330 at 26 fps is roughly 5.6x higher performance and hence is probably above that 5W figure by itself... and I'm still guessing it was running at 6W or above.

 

[Khato]

While I'm aware, that this is a marketing slide, the power draw still seems a bit lower on HPm vs LP, especially at higher frequencies.

I managed to miss that detail in the chart before - thanks. It actually makes sense that they'd realize slightly better efficiency by virtue of HPm apparently being a hkmg process. (TSMC's page actually isn't explicit on this point, but I'd be rather surprised if it isn't.) I'm not certain how much of a gain they'd get out of it though since, as that chart indicates, the primary advantage of moving off the LP process is that you can get more frequency for a given voltage increase.

 

[Exophase]

As for scaling the performance numbers versus Adreno 225, how did you arrive at Adreno 330 only being 4.7x higher performance? The highest GLBenchmark 2.7 score I see for Adreno 225 is 4.6 fps, and the Dell XPS 10 that we obtained the 1W figure for Adreno 225 only gets 3.2 fps. (For which I'd blame it running Windows RT instead of Android.) Even in the best case, Adreno 330 at 26 fps is roughly 5.6x higher performance and hence is probably above that 5W figure by itself... and I'm still guessing it was running at 6W or above.

Arrived there by comparing GLBenchmark 2.5 scores. Shying away from 2.7 because it's more geared for newer GPUs (favors different TEX:ALU ratios). That means I'm also considering the power consumption for GLBench 2.5 instead of 2.7.

I managed to miss that detail in the chart before - thanks. It actually makes sense that they'd realize slightly better efficiency by virtue of HPm apparently being a hkmg process. (TSMC's page actually isn't explicit on this point, but I'd be rather surprised if it isn't.) I'm not certain how much of a gain they'd get out of it though since, as that chart indicates, the primary advantage of moving off the LP process is that you can get more frequency for a given voltage increase.

I did mention HKMG earlier. How would it give more frequency for the same voltage but not also allow lower voltage at the same frequency? The frequency of the GPU part is the same as in Adreno 320, therefore it could be running at a lower voltage now.

 

[Khato]

Arrived there by comparing GLBenchmark 2.5 scores. Shying away from 2.7 because it's more geared for newer GPUs (favors different TEX:ALU ratios). That means I'm also considering the power consumption for GLBench 2.5 instead of 2.7.

Fair enough.

I did mention HKMG earlier. How would it give more frequency for the same voltage but not also allow lower voltage at the same frequency? The frequency of the GPU part is the same as in Adreno 320, therefore it could be running at a lower voltage now.

True, you did. I'd just forgotten about the fact that TSMC's LP wasn't a hkmg process until that chart prompted me to look it up. Sorry about that. Regarding voltage, it is possible, but given the frequency and die area devoted to graphics I'd be surprised if they used high performance transistors for it instead of the low power (more importantly, low leakage) ones. (Going by Chipworks' overview of TSMC's 28nm process variations there, which may well be incorrect.) If still using low power transistors then voltage would likely remain similar, it'd just have better leakage characteristics than the non-hkmg LP process.

 

[Exophase]

I actually think they'd lower dynamic power consumption for the GPU even at the expense of worse leakage. This shouldn't be like CPU cores that will constantly be running little tasks for small time periods. It could spend most of its time outside of games being entirely power gated, especially if there's a separate 2D engine for compositing (and in particular, any time the screen is off). It seems like it'd be better to accept somewhat worse power consumption for very light games if it meant substantially better power consumption for demanding games, but that's just my take on it.