NVIDIA Tegra Architecture

[Exophase]

While it's not surprising that you would assume Tegra 4 would have been in Nexus 7 were it available sooner (and that it's the design win nVidia says they lost) I'm not as sure. Look at the historic progression of Intel's Nexus line:

Nexus One: Qualcomm
Nexus S: Samsung
Galaxy Nexus: TI
Nexus 4: Qualcomm

Nexus 7: nVidia
Nexus 10: Samsung

I don't think it's a coincidence that succeeding Nexus devices have all used different SoC manufacturers. If anything, Tegra 4i showing up in the next Nexus phone would make sense. But not the next 7" tablet.

Also, while I don't think Tegra 4 is out of the question for 7" tablets, I do think S4Pro or S600 is a better fit for a budget tablet with a pretty small battery. There's little doubt in my mind that it'll be more power efficient.

 

[ams]

While it's not surprising that you would assume Tegra 4 would have been in Nexus 7 were it available sooner

That is not what I said. Even if T4 had been available months earlier, it would have been illogical for Google to use it in a Nexus 7 tablet variant considering that it would clearly outperform the flagship Nexus 10 tablet with respect to CPU and GPU performance. And naturally Google will do their best to create products with a variety of different partners over time, so one vendor will never be locked in to any one Nexus branded device.

 

[Exophase]

So what you're saying is that they didn't use Tegra 4 not just because it was delayed but also because it was too good. Got it.

Yeah I don't believe that for a minute. Nexus 10 is an older product and can't keep the performance crown in the Nexus line forever. I'm sure there are already plenty of areas where a Nexus 4 will beat it even if it not in single threaded performance.

 

[ams]

So what you're saying is that they didn't use Tegra 4 not just because it was delayed but also because it was too good. Got it.

Yes, compared to the SoC used in Nexus 10, it is arguably "too good" to be used in a $200 7" tablet. Product positioning matters, and it would make far more sense to use T4 in a Nexus 10 refresh rather than using it in a Nexus 7 refresh (assuming that it would be used in any Nexus tablet device in the first place).

 

[Blazkowicz]

That is not what I said. Even if T4 had been available months earlier, it would have been illogical for Google to use it in a Nexus 7 tablet variant considering that it would clearly outperform the flagship Nexus 10 tablet with respect to CPU and GPU performance. And naturally Google will do their best to create products with a variety of different partners over time, so one vendor will never be locked in to any one Nexus branded device.

Are consumers aware of hardware specs in such a detailed way? I didn't know the Nexus 10 was faster than the Nexus 7 or that one of them is to be considered a flagship.
I guess even litterate people buy based on RAM, flash, display, Android version and not many people know what's a snapdragon 600 or which is better, the qualcomm X or the qualcomm Y.

 

[Ailuros]

On an unrelated note I just noticed that T4 has a quite large amount of custom extensions:

http://gfxbench.com/device.jsp?benchmark=gfx27&D=NVidia+Shield&testgroup=gl

 

[ltcommander.data]

On an unrelated note I just noticed that T4 has a quite large amount of custom extensions:

http://gfxbench.com/device.jsp?benchmark=gfx27&D=NVidia+Shield&testgroup=gl

A lot of those NV extensions seem to correspond to ARB or EXT extensions that are part of OpenGL 3.x or OpenGL ES 3.0 like copy_buffer, draw_instanced, pixel_buffer_object, texture_array, uniform_buffer_object, etc. I guess they thought that making custom versions of standard extensions provides a performance benefit. Developers who make games only for Shield won't mind, if there are going to be any of those, but it would seem to add an extra tweaking step for developers making games targeting multiple GPU vendors.

 

[Ailuros]

First HP Slate 2 results: http://gfxbench.com/device.jsp?benchmark=gfx27&D=HP+Slatebook+10+X2&testgroup=overall

CPU at 1.8GHz and GPU doesn't sound like full 672MHz to me. Since there are no low level results from Shield but there are from the Slatebook, triangle rates are rather "yawnish" on the T4 GPU:

http://gfxbench.com/compare.jsp?D1=HP+Slatebook+10+X2&D2=Asus+Transformer+Pad+TF700T+Infinity&cols=2

 

[Exophase]

Interesting min CPU frequency of 51MHz. That power saver core really dips low.

 

[Helmore]

Roughly the same performance as a Snapdragon Galaxy S4. It's looking like Qualcomm's Snapdragon 800 SoC will be faster than Tegra 4 in pretty much every way.

 

[ams]

Roughly the same performance as a Snapdragon Galaxy S4.

The S600 variant in SGS4 has the same GFXBench 2.7 offscreen (1080p) score of 17.1 average fps as this T4 variant in SlateBook x2, but as I have said before, the result is largely academic because the framerate is not even close to being smooth or playable. The T4 variant in SlateBook x2 does have ~ 14% higher average fps in GFXBench 2.5 offscreen (1080p) than the S600 variant in SGS4, and ~ 24% higher average fps in GFXBench 2.5 offscreen (1080p) than the S600 variant in HTC One. The S800 SoC will obviously help to close the gap in GFXBench 2.5 graphics performance, but the quad-core A15 T4 SoC will still presumably have an advantage in most CPU and browser benchmarks.

 

[Ailuros]

The S600 variant in SGS4 has the same GFXBench 2.7 offscreen (1080p) score of 17.1 average fps as this T4 variant in SlateBook x2, but as I have said before, the result is largely academic because the framerate is not even close to being smooth or playable.

Yes I've read that lame excuse multiple times from you; you obviously still can't comprehend the purpose of a synthetic benchmark. Again results attempt to foresee what each solution could behave in future more demanding applications and it would be purely academic if they wouldn't influence SoC sales at all or any of the few consumers at all.

The T4 variant in SlateBook x2 does have ~ 14% higher average fps in GFXBench 2.5 offscreen (1080p) than the S600 variant in SGS4, and ~ 24% higher average fps in GFXBench 2.5 offscreen (1080p) than the S600 variant in HTC One. The S800 SoC will obviously help to close the gap in GFXBench 2.5 graphics performance, but the quad-core A15 T4 SoC will still presumably have an advantage in most CPU and browser benchmarks.

At what power consumption? Besides here synthetic CPU and browser benchmarks are all of the sudden "relevant" while the majority of them being times more questionable in usefulness than some of the popular GPU benchmarks.

Any leaked data I've seen so far suggests a very good price/performance ratio for S800 solutions. In the meantime bon apetit for you on anyone else to chew on NV's own skewed marketing bullshit and unfortunate speculative material based on past solutions.

 

[ams]

Yes I've read that lame excuse multiple times from you

Lame excuse please It doesn't make sense to have a more forward looking GPU architecture (such as Adreno 320/330) if it doesn't have the GPU horsepower to provide smooth framerates in the first place with more forward looking workloads, and the incredibly low GFXBench 2.7 average fps results on essentially all current or near future ultra mobile devices just reinforces that.

 

[Ailuros]

Lame excuse please It doesn't make sense to have a more forward looking GPU architecture (such as Adreno 320/330) if it doesn't have the GPU horsepower to provide smooth framerates in the first place with more forward looking workloads, and the incredibly low GFXBench 2.7 average fps results on essentially all current or near future ultra mobile devices just reinforces that.

For the moment I can see a tablet design yielding 50 fps average in GLB2.5 compared to a smartphone design which yields 44 fps. Since you have the nasty tendency to bend definitions only in NV's interest let's try to compare a tablet against a tablet or a smartphone vs a smartphone design or do the very same tests with 4xMSAA enabled. The GalaxyS4 will remain playable and the first will be chasing its heels spending the majority of time in single digit frame rates through 2.5.

For the record current and upcoming Adreno 3xx GPUs are as "old" as the ULP GeForce since Tegra1.

 

[ams]

Actually if you look at the median GFXBench scores (which is statistically the score that most SGS4 devices would be centered around), the S600 variant in SGS4 has a GFXBench 2.7 offscreen (1080p) score of 15.3 average fps and a GFXBench 2.5 offscreen (1080p) score of 40.7 average fps. So in comparison to the typical S600 variant in SGS4, the T4 variant in SlateBook x2 has ~ 12% higher average fps in GFXBench 2.7 offscreen (1080p) (which again I consider to be statistically not very relevant right now due to the graphics settings being beyond what this generation of ultra mobile GPU can smoothly handle), and has ~ 23% higher average fps in GFXBench 2.5 offscreen (1080p). As for 4x MSAA, that is about the last thing that is needed right now considering that these high end ultra mobile devices have 4-10" 1080p screens with GPU's that struggle to achieve playable framerates at these resolutions with high detail and no AA (let alone 4x MSAA), so no, the S600 variant of SGS4 will not be playable at high resolutions with high detail and 4x MSAA (and FWIW, T4 should achieve 20-30 average fps offscreen with 4x MSAA enabled in GLB2.5).

 

[Exophase]

As for 4x MSAA, that is about the last thing that is needed right now considering that these high end ultra mobile devices have 4-10" 1080p screens with GPU's that struggle to achieve playable framerates at these resolutions with high detail and no AA (let alone 4x MSAA), so no, the S600 variant of SGS4 will not be playable at high resolutions with high detail and 4x MSAA (and FWIW, T4 should achieve 20-30 average fps offscreen with 4x MSAA enabled in GLB2.5).

You still seem to be having difficulty understanding that a 1080p display doesn't mean that games need to be 1080p.

If you're doing a lower resolution game, and one GPU can add 4x MSAA for little cost while the other can add it for a big cost then yes, it matters. For tilers MSAA can be relatively cheap, which is Ailuros's point.

 

[OlegSH]

You still seem to be having difficulty understanding that a 1080p display doesn't mean that games need to be 1080p.

If you're doing a lower resolution game, and one GPU can add 4x MSAA for little cost while the other can add it for a big cost then yes, it matters. For tilers MSAA can be relatively cheap, which is Ailuros's point.

Actually, MSAA isn't anywhere close to "free" for most tilers, I'm speaking of Adreno and Mali, they are both have huge MSAA hit(much lager relative to desktop IMRs) and I think MSAA is the last thing to implement with high DPI displays, shaders and programmable blending. I've seen a few mobile games with temporall AA - Dragon Slayer on Asus Transformer Prime with Tegra3(which is not supporting MSAA) and Real Racing 3 on IPAD4 in native resolution, TAA is much cheaper, while the IQ is almost the same as with MSAA, thanks to high pixel density displays

 

[Blazkowicz]

1080p on mobile devices?
You might as well run with pixel doubling. 960x540 should give you something to look at.

 

[Deleted member 13524]

1080p on mobile devices?
You might as well run with pixel doubling. 960x540 should give you something to look at.

Exactly.
The Adreno 320 in the S600 should be able to run the T-Rex scene at 540p at over 40 FPS.

The screens support 1080p, but it doesn't mean all games must run at that resolution in a 5" screen using a SoC with single-digit watt power consumption.

The Adreno 320's featureset can be used if the games run at a lower resolution that perfectly fits a 4.7/5" screen.

 

[DSC]

http://www.anandtech.com/show/7014/asus-new-transformer-pad-infinity-tegra-4-2560-x-1600-ips-display

Nice design win for Nvidia and Tegra 4.