NVIDIA Tegra Architecture

[Mariner]

Checking the linked GSIII benchmarks (notably the GLBenchmark offscreen results), it would appear that the Mali MP in the Exynos runs at the same speed as the RK3066.

Interesting. I'd have thought the Exynos might have been clocked a little higher as it is built on a smaller process than RK3066. I suppose power constraints are more important for a chip which is going to be used in a phone, even one with a 4.8" screen!

You're right, though. If the T3+ does have a 30% boost in GPU performance, it ought to be within 10% of the Mali MP. Still a little disappointing.

Wait a minute. What am I saying? I own a few hundred quid of shares of ImgTch so should be pleased the Tegra series aren't graphical behemoths.

 

[metafor]

I never said that. I said that Tegra 3+ (aka T37, which is supposed to have at least 20-30% faster CPU/GPU operating frequencies vs. T30L, and probably updated drivers too) should be reasonably competitive with the Mali-400/MP4 that is used in the Galaxy S III international phone.

It's interesting that stuff like this pops up around.

 

[ams]

You're right, though. If the T3+ does have a 30% boost in GPU performance, it ought to be within 10% of the Mali MP. Still a little disappointing.

Sure, I hear you. Fortunately for NVIDIA is that most tablets used with Tegra 3 target a screen resolution that is typically no higher than 720p HD. For 1080p full HD tablets, it would be wise to wait for Tegra 4.

Wait a minute. What am I saying? I own a few hundred quid of shares of ImgTch so should be pleased the Tegra series aren't graphical behemoths.

ImgTch is a great company. Same with ARMH, NVDA, AMD, etc. You should seriously consider investing in all these companies (the latter three being heavily discounted right now relative to their 52 week high) to broaden your portfolio. Hold the shares for at least one or two years, and you should be in really good shape

 

[Lazy8s]

What I was stating with regard to Tegra 2/3 and A5/A5X is that I believe the Tegra architecture wouldn't have the efficiency to match their respective performance even if more silicon were added to the Tegras (fabbed in a comparable manufacturing process) since power and heat would become too limiting to be practical.

Matching A5 and A5X's graphics performance isn't just about devoting a lot of silicon to the GPU. Samsung's Exynos 4210 was similar in size to the A5, yet the differences in the designs of their respective GPUs contributed to a gap in graphics performance between them.

 

[ams]

I don't think there is any evidence to indicate that power and heat would become too limiting if Tegra 3 used a larger SoC die size and devoted more of that area to the GPU. Tegra 3 easily has low enough power consumption (typically less than 1w?) to be used in a smartphone. In a tablet/clamshell device, power consumption doesn't need to be anywhere near that low.

Naturally improved graphics performance relative to other GPU architectures is not simply about devoting more silicon to the GPU. But for any given GPU architecture, being able to use a larger GPU die size vs. a smaller GPU die size will almost always bring about performance improvements. That is why a GTX 580 will handily outperform a GTX 560, why a GTX 480 will handily outperform a GTX 460, etc. etc.

 

[Ailuros]

I don't think there is any evidence to indicate that power and heat would become too limiting if Tegra 3 used a larger SoC die size and devoted more of that area to the GPU. Tegra 3 easily has low enough power consumption (typically less than 1w?) to be used in a smartphone. In a tablet/clamshell device, power consumption doesn't need to be anywhere near that low.

The ULP GeForce in T30 has a peak frequency of 520MHz and for AP30 of 416MHz. Theoretically if NV would had devoted more die area to the GPU block, they might had not needed as high frequencies under 40nm/TSMC and power consumption might had balanced out somewhat, within specific boundaries.

The SGX543MP2 in A5 and SGX543MP4 in A5X should be clocked at 250MHz each.

Naturally improved graphics performance relative to other GPU architectures is not simply about devoting more silicon to the GPU. But for any given GPU architecture, being able to use a larger GPU die size vs. a smaller GPU die size will almost always bring about performance improvements. That is why a GTX 580 will handily outperform a GTX 560, why a GTX 480 will handily outperform a GTX 460, etc. etc.

Yes but any of those performance improvements aren't for free either in terms of power consumption. Take your examples of desktop GPUs above one by one and look at their respective power consumptions/TDP. A GTX580 has a TDP of 244W and is roughly >40% faster on average than a 170W TDP 560Ti and no it's obviously not a coincidence.

 

[Lazy8s]

Yes, clock speed of the Tegra GPU was the main indicator I cited originally for my belief of the architecture having less headroom to scale up. For context (while not directly applicable), the Exynos 4210 at a similar die size to the A5 clocked its Mali-400 MP4 around 267 MHz on 45nm (and 440+ MHz on 32nm).

 

[ams]

Yes, clock speed of the Tegra GPU was the main indicator I cited originally for my belief of the architecture having less headroom to scale up. For context (while not directly applicable), the Exynos 4210 at a similar die size to the A5 clocked its Mali-400 MP4 around 267 MHz on 45nm (and 440+ MHz on 32nm).

As Ailuros pointed out, if NVIDIA was in a position to use a much larger GPU die size in order to dramatically increase the number of execution units, then GPU operating frequency could have been much lower. The Exynos 4210 SoC in fact uses a much larger die size than Tegra 3 SoC.

Yes but any of those performance improvements aren't for free either in terms of power consumption.

True, but Tegra 3 has a typical power consumption of only ~1w, and power consumption in a tablet/clamshell-type device doesn't need to be anywhere near that low.

 

[Laurent06]

True, but Tegra 3 has a typical power consumption of only ~1w, and power consumption in a tablet/clamshell-type device doesn't need to be anywhere near that low.

What does typical mean here? I'd be surprised, if with all 4 cores running NEON code with GPU crunching, the SoC could be consuming only 1W.

 

[Lazy8s]

Especially for the tablet variant of a mobile SoC, spikes well above 1W under load aren't uncommon, yet draw should hopefully keep to a half-Watt or less for much of the processing. Any successful mobile SoC would meet those demands of the form factor, but I don't see where Tegra stands out in that regard compared to the other successful SoCs on the market.

My point with the Exynos 4210 example was to show that another manufacturer did strike a balance where the GPU clock speed could be lowered and the amount of GPU silicon was raised for execution resources, yet underlying architectures still played a role in a significant difference in graphics performance versus the A5.

 

[metafor]

I don't think there is any evidence to indicate that power and heat would become too limiting if Tegra 3 used a larger SoC die size and devoted more of that area to the GPU. Tegra 3 easily has low enough power consumption (typically less than 1w?) to be used in a smartphone. In a tablet/clamshell device, power consumption doesn't need to be anywhere near that low.

"typical" is a tricky number. At least based on measurements of MSM8960, the GPU alone can run close to 2W while in usage. Obviously "typical use" doesn't include heavy gaming for marketing.

When all of the components of the SoC are in use, I wouldn't be surprised to see 3-5W being consumed.

 

[Ailuros]

As Ailuros pointed out, if NVIDIA was in a position to use a much larger GPU die size in order to dramatically increase the number of execution units, then GPU operating frequency could have been much lower. The Exynos 4210 SoC in fact uses a much larger die size than Tegra 3 SoC.

The ULP GeForce didn't undergo any major changes from Tegra1 to Tegra3. Tegra2 just got about twice the frequency and they've added another PS Vec4 ALU to the ULP GF in T3 amongst some other minor refinements. Chances are high that rumors circulating behind the curtains that NV had a miniscule of engineering resources devoted to the Tegra team up to T3 make more sense than anything else.

If true it's a miracle how they managed to pull even the changes listed above and of course even more so still staying competitive overall from the GPU side of things ('blame' sw/compiler experience and resources amongst others).

Here I'd expect a quite sizeable change in terms of GPU performance starting with Wayne/Tegra4, also because background rumors whisper of a sizeable amount of resources that have been re-devoted to the Tegra GPU team quite some time ago.

True, but Tegra 3 has a typical power consumption of only ~1w, and power consumption in a tablet/clamshell-type device doesn't need to be anywhere near that low.

I think adding on top of all the other replies above is redundant. Overall I don't have a single problem claiming that Tegra3 is quite a well positioned SoC overall and no none of them can win everywhere or have it all.

 

[anexanhume]

I would hope that Wayne has something new in the GPU area. Rogue is going to be a big step up when it hits. However, I also haven't seen any indication GPU strength has anything to do with positioning one handset against another (unlike consoles and handhelds), so we'll see if it even matters.

 

[Ailuros]

I would hope that Wayne has something new in the GPU area. Rogue is going to be a big step up when it hits. However, I also haven't seen any indication GPU strength has anything to do with positioning one handset against another (unlike consoles and handhelds), so we'll see if it even matters.

It's less about hw itself and if the end user can make any use of additional functionalities and more a further push into the heterogeneous computing direction for small form factor SoCs if the Wayne GPU should contain USC ALUs as I assume.

 

[anexanhume]

It's less about hw itself and if the end user can make any use of additional functionalities and more a further push into the heterogeneous computing direction for small form factor SoCs if the Wayne GPU should contain USC ALUs as I assume.

Genuine question. Do you see the commitment to heterogeneous computing at google for android? I honestly don't know. Only apple has made their ambitions clear there.

 

[Ailuros]

Genuine question. Do you see the commitment to heterogeneous computing at google for android? I honestly don't know. Only apple has made their ambitions clear there.

Is there any GPU vendor for the small form factor market that doesn't have an OpenCL capable GPU planned for the next generation?

I personally see GPGPU as a general necessity to squeeze more efficiency out of an SoC, especially since you have in a SoC a multitude of units fighting for the very same bandwidth.

http://withimagination.imgtec.com/?p=576

Patrick Moorhead, principal analyst at Moor Insights & Strategy said: “This [foundation] will place CPUs, GPUs, and other accelerators as equal citizens across computing platforms. If the HSA Foundation gets support from Microsoft with Visual Studio, Google with Android SDK, and Apple with XCode, this would be a real game changer.”

I'd be very surprised if NV doesn't enter the HSA foundation amongst others over time.

 

[rpg.314]

Is there any GPU vendor for the small form factor market that doesn't have an OpenCL capable GPU planned for the next generation?

I personally see GPGPU as a general necessity to squeeze more efficiency out of an SoC, especially since you have in a SoC a multitude of units fighting for the very same bandwidth.

http://withimagination.imgtec.com/?p=576



I'd be very surprised if NV doesn't enter the HSA foundation amongst others over time.

Depends on whether Apple/Google/MS go for HSA foundation or go with some other consortium. I don't see any of them deciding to do their own roadmaps, even Apple.

 

[Ailuros]

Depends on whether Apple/Google/MS go for HSA foundation or go with some other consortium. I don't see any of them deciding to do their own roadmaps, even Apple.

There's a reason I highlighted the "if" in the quote above. I don't think the question was whether each vendor will bank with A, B or C foundation/initiative, but whether there's any sort of interest for GPGPU from Google's side for Android. I obviously don't have a clear answer for the latter, but I'd be also quite surprised if it wouldn't interest Google at all.

 

[Deleted member 13524]

Has anyone tried Splashtop THD for Tegra 3 devices?

It's supposedly using the same system that Jen Hsu Huang showed in the "Virtual GPUs" presentation for streaming games from the PC into the tablet.

Getting stuff to work in fullscreen is still a bit rough around the edges, it could become a nice attempt to bring a portable gaming PC to the living room, like a Wii U.

 

[Jubei]

“Expect our upcoming Wayne chip next year to provide double the Tegra 3 performance, and Stark a few years from now should provide up to twenty five times the Tegra 3 performance," said Stam.

http://www.fudzilla.com/home/item/28159-nvidia-hopes-to-make-good-on-icera-acquisition