NVIDIA Tegra Architecture

[OlegSH]

NV never claimed 850 or 950 MHz in a thin tablet, but lots of folk on the internet claimed the Lenovo 28" television represented the pinnacle of K1 performance. Interesting to see that it's clocked higher in a thin tablet than in a television.

I think it's a common mistake to think that something big(like TV) should include the best possible SOC SKU since it's BIG, while in practice most of smart TVs are built on low end SKUs because there is no demand for high end SOCs in such devices for different reasons, it's a miracle that Lenovo decided to go Tegra(even if it's a low end SKU) instead of cheap low perf Mediatek SKU. I believe Xiaomi Mi Pad is based on high end SKU for tablets since it's Manhattan offscreen result is very close to Jetson TK1 one(30 vs 32 fps) and Jetson TK1 has 850 Mhz max GPU frequency

 

[OlegSH]

Don't tell me that you'd rather prefer the MiPad to yield 30 fps at 850 instead of 650MHz ROFL

Actually I'd rather prefer the MiPad to yield 30 fps at 850 instead of 650MHz in Manhattan test because it uses deferred rendering, so it should be bandwidth and shader bound, so the test can be bandwidth bound in the end on immediate architecture like Kepler, which would mean that Tegra K1 can be more than 2.3 times faster than A7's Rogue in shader bound forward rendering apps

 

[Ailuros]

Actually I'd rather prefer the MiPad to yield 30 fps at 850 instead of 650MHz in Manhattan test because it uses deferred rendering, so it should be bandwidth and shader bound, so the test can be bandwidth bound in the end on immediate architecture like Kepler, which would mean that Tegra K1 can be more than 2.3 times faster than A7's Rogue in shader bound forward rendering apps

Manhattan is mostly ALU bound and it shows in the majority of all platform results. Since both the MiPad as the iPad mini retina have a native 2048*1536 native resolution you can always compare onscreen results.

Other than that if you look at the onscreen fill rate result of the Lenovo at 850MHz the fill rate efficiency would me more than low.

 

[OlegSH]

Other than that if you look at the onscreen fill rate result of the Lenovo at 850MHz the fill rate efficiency would me more than low.

There is clearly a lower clocked SKU in Lenovo compared to MiPad, CPU alone is downclocked by 10%

Manhattan is mostly ALU bound and it shows in the majority of all platform results.

But still it's a deffered with mediocre geometry complexity and should benefit from bandwidth savings of tiled architectures, while tegra scaling with frequency could be below lenear because of constant bandwidth

 

[DSC]

http://arstechnica.com/gadgets/2014/05/nvidia-inside-hands-on-with-audi-lamborghini-and-tesla/

 

[ams]

Just out of curiosity where does it anywhere state that the 804MHz are for the GPU frequency?

My bad, you are right, that is clearly not the GPU frequency. In that case, I have no idea what frequency the GPU was operating at. But if Oleg is right that Jetson TK1 is only 2fps faster than Mi Pad in GFX Bench 3.0 Offscreen, then doesn't that imply that the max GPU frequency is relatively high in Mi Pad?

 

[Ailuros]

But still it's a deffered with mediocre geometry complexity and should benefit from bandwidth savings of tiled architectures, while tegra scaling with frequency could be below lenear because of constant bandwidth

There's always a gap between "should" and "is". The difference between the A7 GPU and the MiPad GPU shrinks somewhat when you go up to both their native resolutions but not by any big margin, probably because Manhattan is as ALU bound and the A7 GPU can't really do any wonders with its 115 GFLOPs FP32 peak performance.

http://gfxbench.com/compare.jsp?ben...roid&api1=gl&D2=Apple+iPad+Mini+Retina&cols=2

Onscreen between those two at 2048*1536 the difference is at 1.9x, while at 1080p offscreen at 2.3x.

 

[Ailuros]

My bad, you are right, that is clearly not the GPU frequency. In that case, I have no idea what frequency the GPU was operating at. But if Oleg is right that Jetson TK1 is only 2fps faster than Mi Pad in GFX Bench 3.0 Offscreen, then doesn't that imply that the max GPU frequency is relatively high in Mi Pad?

NV states in the latest Jetson whitepaper that its ~2.5x times faster than the A7 GPU at 13 fps. That's either 32,5 or 33,0 fps for the Jetson and while in essence you're right (no one sane will bargain whether it's 2, 2.5 or 3.5 fps difference) Lord knows if they have optimized the driver even further.

From what I've been told the GK20A fillrate efficiency is outstanding in GFXbench3.0 but I can't of course guarantee it in any way.

http://gfxbench.com/device.jsp?benchmark=gfx30&os=Android&api=gl&D=Lenovo ThinkVision 28

Onscreen 2048*1536 = 3799 / 8 TMUs = 475MHz
Offscreen 1920*1080 = 4644 / 8 TMUs = 581MHz

The latter fits better into the same picture as the supposed info tidbit above. At 650 peak frequency you have a ~90% fillrate efficiency and at 850 peak a ~70% efficiency.

 

[DSC]

http://vr-zone.com/articles/nvidias-tegra-k1-snapdragon-killer/77762.html

 

[Deleted member 13524]

Now I'm wishing that nVidia had gotten their hands into a x86 license and this would've been a chip to go directly against Mullins and BayTrail.

Maybe Intel will license nVidia's iGPUs for a future Atom SoC?

 

[Alexko]

That doesn't tell us much about power use, but more to the point, TK1 will have to compete against Qualcomm's S805, not S801, which is just a minor update of the S800.

 

[lanek]

I dont think they will cancel all their roadmaps right now, but.... it seems they will drop tablet phone market, and concentrate on embedded ( cars etc ) and gaming ( shield etc. )

http://www.cnet.com/news/nvidia-ceo-sees-future-in-cars-and-gaming-q-a/

I copy paste the short version from guru3D:

Nvidia learned the hard way that the mobile telephone and tablet market is harsh, brutal and monopolized. In an interview with CNET NVIDIA CEO Jen-Hsun Huang pretty much dropped a bomb when I learned that Nvidia is no longer targeting smartphones and tablets with Tegra. They are no longer focussing on that that market.
Nvidia can't seem to gain ground with their solution and as such will shift their focus with this SoC on computing and gaming. Huang explains it's not NVIDIA's goal to go after commodity or mainstream devices, they want to be in performance-oriented, visual computing-oriented, gaming-oriented devices. Here's a part of the interview:


Does that mean we're not going to see Nvidia in smartphones? You earlier said gaining share in smartphones was a matter of releasing a processor, called Tegra 4i, that integrated 4G with your apps processor. Has that not worked?
Huang: [Tegra 4i] wasn't that successful for us. I would say that when we first started this, we thought that bringing 4G to entry-level phones, mainstream phones, integrated with our apps processor would be a real opportunity. I think that the phone marketplace has commoditized really, really fast. It is not our strategy to go after commodity phones. It is not our strategy to go after mainstream devices. But our strategy is to focus on performance-oriented, visual computing-oriented, gaming-oriented devices where we can add a lot of value.


Why did Tegra struggle in smartphones?
Huang: Our focus as a company is still performance-oriented. The mainstream phone market commoditized so fast that really the...differentiators were price. And you can see the pressure that MediaTek is putting on Qualcomm, and you can see the pressure that MediaTek is putting on Marvell and Broadcom and all of these companies. Because guess what? They're the lowest-cost provider. I think that for mainstream phones, there's one strategy that really works right now, which is price. That's not our differentiator. That's not what we do for a living.


Will there be a Tegra 5i or whatever you would call a chip that integrates 4G with an apps processor?
Huang: We don't talk about future products, but I also haven't talked about T5i.


You delayed Tegra 4 for Tegra 4i. Did that turn out to be a mistake? Did you miss this whole design cycle?
Huang: I would say that Tegra 4i didn't pan out. We learned a lot in the process. But there are many things in our company that didn't pan out. That's OK. If you want to be an innovative company, you have to fail.
Look, we built a great chip. LG's shipping it in the rest of the world outside the United States. It's a fantastic processor. But from a business strategy, it wasn't a success. So I learned a lot from it. It's OK. I'm glad I did it, and now we're moving on.

 

[Deleted member 13524]

Wow, they show no hesitation in calling T4i a business failure.

I guess they really had a very small window of opportunity of building a solid mainstream ground with a LTE chip before Mediatek starts blasting every competitor with their LTE big.LITTLE solutions.

 

[lanek]

Wow, they show no hesitation in calling T4i a business failure.

I guess they really had a very small window of opportunity of building a solid mainstream ground with a LTE chip before Mediatek starts blasting every competitor with their LTE big.LITTLE solutions.

Its an hard competition market, i think even without Mediatek, they had not really much luck against qualcomm allready... But i like the honesty we find in his responses..

 

[Helmore]

That doesn't tell us much about power use, but more to the point, TK1 will have to compete against Qualcomm's S805, not S801, which is just a minor update of the S800.

True, but I personally think that the S805 won't be as fast as the K1. The question is, which one will be better when it comes to power use? There I think that the S805 will be better.

 

[Alexko]

True, but I personally think that the S805 won't be as fast as the K1. The question is, which one will be better when it comes to power use? There I think that the S805 will be better.

We'll see. We should also bear in mind that the S810 (with Adreno 430) is expected about 5 months after the S805, which means it should compete with TK1-Denver, whose GPU is the same as TK1-A15's.

 

[Deleted member 13524]

We'll see. We should also bear in mind that the S810 (with Adreno 430) is expected about 5 months after the S805, which means it should compete with TK1-Denver, whose GPU is the same as TK1-A15's.

I wonder if TK1-Denver will ever be a final consumer product and not just an internal solution to prepare and learn for Erista or "TM-1" (M for Maxwell).

It seems to me that the S810 will arrive awfully close to Erista.

 

[wco81]

So they weren't willing to compete on price, which is mainly why they didn't get design wins.

But when he talks about performance oriented products, could it also be code for the fact that the Tegra 4s were not power-efficient enough relative to the competition?

I don't think the car market is as huge as he thinks. Sure people will buy some entertainment package with their luxury cars (sometimes there isn't a choice) but more people would prefer that cars just offer dumb screens for their mobile devices to stream to.

And if they're not going to bother trying in the mobile device market at all …

Phones and tablets should have much greater volumes than car systems.

 

[Ailuros]

So they weren't willing to compete on price, which is mainly why they didn't get design wins.

But when he talks about performance oriented products, could it also be code for the fact that the Tegra 4s were not power-efficient enough relative to the competition?

I don't think the car market is as huge as he thinks. Sure people will buy some entertainment package with their luxury cars (sometimes there isn't a choice) but more people would prefer that cars just offer dumb screens for their mobile devices to stream to.

And if they're not going to bother trying in the mobile device market at all …

Phones and tablets should have much greater volumes than car systems.

Yes and no; however it might be just me but that's one of the very honest interviews I've read from JHH in a long time.

 

[lanek]

Yes and no; however it might be just me but that's one of the very honest interviews I've read from JHH in a long time.

Same feeling, and for all the comment i have read, i, we, are not alone to think this .