NVIDIA Tegra Architecture

That's funny, why then Vizio demoed a T4 tab at CES, I am pretty sure Asus will show one at MWC too

 

No problem. By the way, as mentioned earlier, I do agree with Fudo that the Snapdragon S4 Pro would be a good choice for a Nexus 7.7, positioned above the Nexus 7 and positioned below the Nexus 10. Tegra 4 is not even ready yet, and would be out of place in the current Nexus lineup if used in a 7.7 tablet variant.

 

Nexus4 and iPhone5 appeared on shelves more or less at the same timeframe. Both side's drivers should be more or less of the same age and since I can sense the reply already you better should double check if the iPhone5 GPU driver is unique or the same between a number of devices.

No matter what you want to consider it, iPhones are one single device selling with an insane profit margin, exactly because it's as expensive as it is. The Nexus4 sold something 1Mio devices up to now afaik; how many iPhone5 sets since it's introduction? It might not sound relevant to the above, but:

1.) Performance per se is not always the most defining factor as long as the difference is as close at it is in that case.

2.) The Adreno320 is by a margin faster in GL2.5, it's not in 2.1, not in effective fillrates, not in geometry throughput etc. according to the results here: http://www.glbenchmark.com/compare.jsp?D1=Apple+iPhone+5&D2=Google+Nexus+4&cols=2

3.) Are all Adreno320 smartphones performing on the same level as the Nexus4? I don't think so: http://www.glbenchmark.com/compare.jsp?D1=HTC+Butterfly&D2=Apple+iPhone+5&cols=2 How about about those Adreno320 cases where the performance is even way below 3k frames?

After some point it starts getting silly debating the same things all over again and despite that am I documenting my claims you're not willing to move a single inch from your conception. Now I'll lean out of the window and claim that Apple could if they wanted to to surpass through perfectly legitimate driver optimisations by a healthy margin even the Nexus4. What is not guaranteed in such a case is that the iPhone5 will have the same GPU and in extension device stability as it has right now.

Given that the Nexus4 poses on top of all 320 powered smartphones but comes with thermal throttling problems (which as I noted I'm not willing to generalize over all S4/Adreno320 device implementations until further data is known) but if a serious website like Anand mentions "freezer" in its benchmark results for the Nexus4 and has mentioned in detail a throttling problem that affects that particular device (and probably to some extend the LG Optimus G) I wouldn't get so stuck in your place on that result, but would try to think of the most stable device/platform behaviour with a Adreno320 GPU and then make performance comparisons.

 

8 cluster Mali450 is OGL_ES2.0; practically just a Mali400MP4 times 2.

Developers can have fun even on that one developing mobile games for the future which likely will played on such devices in extreme low quantities in the future and probably at equally crapalicious performance as on today's Adreno320 smartphones.

All of us learn something every day

 

Looking at the best results on glbenchmark is a bit misleading IMO. Nexus 4 scores 33,7 fps at its best but the average result drops down to 26 fps

whereas the iphone 5 scores 29 fps but the average score is still very close at 28 fps

And comparing to the other Adreno 320 phones that are all scoring somewhere between 29-31 fps, it puts Adreno 320 neck to neck with SGX543MP3. The difference is that the latter is clocked much lower, we will see what happens when Samsung releases series 5 chip that is clocked closer to the competition

 

According to the leaked Qualcomm slide (http://forum.beyond3d.com/showpost.php?p=1700375&postcount=308), the Adreno 320 GPU used in the S4 Pro SoC has a GPU clock operating frequency of 400 MHz. The SGX 543MP3 GPU used in the A6 SoC should have a GPU clock operating frequency of at least 325 MHz (and possibly even higher than that). So the differences in clock speed between these two GPU's is significant but not so huge.

 

Hmm i must have missed this. I was under the assumption that it was clocked at 267 MHz based on Anandtechs preview

 

Yeah, I think that Anandtech's guess in the iphone 5 preview was wrong. The GLBenchmark 2.5 data on iphone 5 vs. ipad 3, the GFLOP throughput comparison on iphone 5 vs. ipad 3, the CPU-to-GPU clock ratio comparison in past Apple i-devices, all point to a GPU clock operating frequency that is at least 325MHz for the iphone 5 GPU.

 

You really think that ~30fps on GLBench2.5 isn't struggling?

Not wanting to dwell on this, but you did quote ARM marketing bullshit about GPU CPU integration...

ALL games or just "generic" mobile content that is written for the lowest common (performance) denominator?

Yes agree, and the first thing you need for this, before throwing esoteric features at the problem, is performance, this is particularly true within an extremely limited power budget.

 

Wait, are you saying that current mobile games don't scale graphics settings based on the hardware that's available? Just because they don't reveal those settings to the user doesn't mean they aren't present and being used by the developer to optimize for each device. GTA III/VC enables higher polygon models, textures, lighting effects on A5/A6 Apple devices compared to A4 devices. The Android version, of course, gets a full user configurable graphics settings menu. Infinity Blade I/II optimizes different combinations of overall display resolution, texture quality, anti-aliasing, and shaders to make the most out of each supported device. Real Racing 2, NOVA 3, Modern Combat 4, etc similarly have differing graphical quality depending on device.

For games where graphics are important and have the development budget for it, developers have been able to scale to take advantage of both high-end and low-end devices. I just don't see Infinity Blade II, for example, as having left a lot of performance on the table on A5 devices, which had the fastest GPU available when the game was in development, even though it also supports the much weaker iPhone 3GS GPU. If the A5 was being underutilized, accommodating the Retina iPad wouldn't have been such an issue for games. Admittedly, developers hard optimizing settings for each device is easier for iOS where there are fewer devices than Android where a few generalized performance classes are probably used. User configurable options would make sense for Android so users can further tweak things for their specific device when the developer hasn't been able to.

 

Up to 40% thermal throttling isn't something to sneeze over: http://www.anandtech.com/show/6440/google-nexus-4-review/3 In GL2.5 offscreen 1080p it's at 18.5 fps vs. 30.5 fps in the "freezer". Now am I not jumping to any conclusions yet because the issue hasn't been revisited ever since and yes crap happens even in best families, but with such details at hand I'd be a bit more cautious in your place with performance claims, especially if anyone can run singled out tests in GLB and submit them. The real question would be how it looks like in real world mobile games and first above all if the problem has been rectified on Nexus4 devices in the meantime.

According to Qualcomm's own roadmap it sounds like the same amount of units for Adreno330 vs. 320 (at least it's the same amount of TMUs for sure - 8 for both). Given that according to that data the 330 is clocked at 450MHz and the 320 at 400MHz and they claim the 320 is 23x Adreno200, while 330= 30x Adreno330, there's a 30% claimed difference between the two. The frequency difference is at 11%, meaning that assuming perfect performance scaling with frequency gives you another 19% which would come hypothetically from some hw improvements + obviously quite a bit higher SoC bandwidth. If there are driver/compiler improvements down the line (more than likely) those could under normal conditions affect not only 3xx class Adrenos, meaning that in such a case the 30% claimed performance difference between 320 and 330 won't change much if at all.

 

http://withimagination.imgtec.com/i...taking-texture-compression-to-a-new-dimension

One aspect of how well Series 5XT fairs against the Adreno 320 is quality of texture formats. Imagination seems quite confident that PVRTC2 provides superior quality/size compared to ETC2/EAC and other image formats. The Amazon Fire HD 8.9 has PVRTC2 driver support although I don't know if any games use it yet. I'm not sure if ETC2 is in shipping devices drivers yet as an OGL ES 2.0 extension while we wait for OGL ES 3.0 drivers.

 

To bounce back a wee bit more on topic: if Fudo is right and NV has been set back with T4 hw due to a spin for two quarters I truly wonder if their expectation of having the fastest tablet SoC for this year will hold up to reality.

 

Firstly, at no point have I said there is performance to spare on current hardware, that's your claim. I have specifically stated that current generation mobile GPU's stuggle with _demanding_ content, e.g. ~30fps on glbench2.5 IS struggling, not that GLB2.5 is particularly strenuous by modern standards.

Secondly I have specifically equated performance to IQ i.e. more peformance = better IQ, at no point have I said we need performance before IQ.

I think the key difference is that you seem to have exclusely couple IQ to features, I'm pointing out that IQ is more related to performance at this time.

No. We agree that titles written for lowest common denominator could have IQ improvements if optimimised for higher performance platforms. However those IQ improvements will still primarily be capped by performance not features.

No, I've repeatedly pointed out that old hardware does not have the scope to run _demanding_ content, as is clearly illustrated by GLBench2.5. However yes, the disparity between low and high end is likely to grow in the near term and of course, adding extra features at the high end would only make that disparity worse.

Obviously.

 

John, I think french toast needs a real-world example where the limitation is clearly performance rather than features. Here's the most perfect example I can think of: The Witcher 2. It's still fully a DX9 game (which looks better than many DX10 and even DX11 games). SGX544/SGX554 are fully DX9 compliant.

Therefore, the only limitation for SGX544 to run The Witcher 2 at maximum quality settings is performance and nothing else. Now tell me honestly that this doesn't look many generations ahead of anything you've ever seen on a handheld:

http://assets1.ignimgs.com/2012/03/...nhanced-edition-20120308095456731-3611513.jpg
http://www.geforce.com/Active/en_US...s/TheWitcher2AssassinsOfKings-05042011-06.jpg
http://www.geforce.com/Active/en_US...s/TheWitcher2AssassinsOfKings-05042011-03.jpg
http://www.geforce.com/Active/en_US...s/TheWitcher2AssassinsOfKings-05042011-08.jpg

As I said, there's nothing that prevents SGX544 from running this except performance. Now that's not quite true for OpenGL ES2.0 in general; you do need MRT support for deferred rendering among many other things, but SGX will support all of these features and Khronos is widely expected to standardise a significant number of OpenGL ES3.0 features into standard extensions that multiple vendors will support.

Yes, there are features in DX10/DX11 (some of which are in ES3; many of which are not) which would allow The Witcher 2 to run faster/more efficiently on the same hardware. But they also cost area/power to implement so the increase in perf/watt might not be as large as you'd expect it to be. And I would describe ES3 as a very evolutionary API, its main advantage is that it standardises many extensions that have existed on some hardware for many years.

If you look at Battlefield 3, it makes very good use of tesselation, but there's only one real reason why it's a DX11-only game: DirectCompute shaders which they use for their excellent deferred rendering implementation. OpenGL 4.2 now includes compute shaders inside the API rather than requiring the developer to use OpenCL (and all the problems that can cause) but unfortunately that's not the case in OGL ES3.0 yet. Personally I completely agree with the idea that graphics is the killer app for compute, but unfortunately handhelds are not ready for it either in terms of API support or architecture (too many incompatible optimisations for different architectures e.g. shared/local memory bank conflicts). That's a much bigger problem than ES2 vs ES3 support in my opinion.

P.S.: On a separate note, please let me know when any handheld game is anywhere as awesome as TW2. Except for some adventure game ports, I haven't seen any good story-based game, and nothing as immersive (possibly because handheld games are more optimised towards shorter gameplay sessions).

 

Battlefield 3 does support DX10 GPUs. I thought therefore DirectCompute was optional rather than falling back to Compute Shader 4.x?

I believe compute shaders were integrated in OpenGL 4.3 not 4.2.

Besides graphical/performance limitations and the naturally shorter gameplay sessions of mobile the other issue is going to be development budget impeding development of very immersive mobile games. I don't think even big developers are generally willing to risk the 10s or 100s of millions of dollars and "when it's done" indefinite development time that console/PC games can get on a mobile game. Admittedly big budget doesn't necessarily translate into a fun, immersive game, but mobile games are certainly not on an even playing field in this regard.