Tegra 3 officially announced; in tablets by August, smartphones by Christmas

[Mike11]

Tegra 3 officially announced; in tablets by August 2011, smartphones by the end of 2011

Project Kal-El:

• World's first mobile quad-core CPU
• New 12-core NVIDIA GPU, with support for 3D stereo
• Extreme HD – 2560 x 1600
• 5x Tegra 2

Nvidia Blog:
http://blogs.nvidia.com/2011/02/teg...-chip-worlds-first-quadcore-mobile-processor/

And further down the road: 75x Tegra 2 performance in 2014.

[...] Today at Mobile World Congress (MWC), we demonstrated this little beauty running in an Android tablet. We not only showed that it was alive. We showed it browsing the Web, running games and streaming amazing video. This wasn’t your average amazing video. It was 1440p video content running on a 2560×1600 panel. That will enable mobile devices to output to the highest resolution monitors or tablets equipped with a 10.1-inch display with 300 DPI.
I should mention, in passing, that this is the first mobile quad-core processor. And that it contains a new 12 core GeForce GPU. Our customers are getting samples now, and they’re planning production in August.

While demonstrating Project Kal-El was exciting, we also gave a glimpse of our roadmap here at MWC. It includes Projects codenamed Wayne, Logan, and Stark, coming out in a steady one-year cadence over the next three years. You might well ask, What on earth can be done with nearly 75x improvement in performance over Tegra 2 that Stark will provide in 2014? Our customers and partners have already indicated that they’re confident they can use everything we give them. [...]

http://www.engadget.com/2011/02/15/nvidia-announces-quad-core-kal-el-soc-promises-it-in-tablets-by/

During the session, Carmack said that, despite its added power, Kal-El will actually be more power efficient than Tegra 2 is today. He said that, while playing that 1440p video in the demo, the CPU and chipset only consumed 400 milliwatts of power.

http://blog.laptopmag.com/nvidia-de...power-of-tegra-2-but-with-longer-battery-life

Wow, August is a lot earlier than I expected. And that roadmap looks quite nice too.

 

[rpg.314]

Any idea whether denver is logan or stark?

 

[Exophase]

Hm, so the GPU is now 12-core now instead of.. what, 8-core for Tegra 2? I never really got what that actually meant. I wish nVidia would talk in more discrete real life terms instead of just made up marketing terms. I wonder what this actually means though. That's only 1.5x the cores nVidia, how much better can it be ;p

It's clear from the Coremark scores that they have no problem pulling numbers straight from the website that were compiled with a much older version of GCC. It's good that they at least mentioned it on the slide, since it's such an obvious gotcha. One interesting thing is that all of the Tegra 2 Coremark submissions get about the same amount, while the Intel ones fluctuate quite a lot. Coremark scores are kind of hard to take seriously all around, but I guess that's not really going to change.

 

[Squilliam]

How exactly are they going to deliver 75* more performance without first turning Tegra into a laptop or desktop computing platform?

 

[Erinyes]

Seriously impressive that they're managed to get it up and running in just 12 days! They demoed it running flawlessly on A0 silicon. But i was really surprised to read it is on 40 nm. I was expecting them to go with 28nm. So that means Wayne is going to quad Cortex A15 on 28 nm.

But i think Nvidia is missing a trick here, they should also be working on a dual core Cortex A15 implementation. I dont see quad cores taking off in smartphones so quickly. Of course there are plenty of other markets where the quad core Tegra 3 will find application, maybe they chose to concentrate on tablets and other devices rather than smartphones.

Any idea whether denver is logan or stark?

Good question. From the roadmap, given the huge jump in performance from Wayne to Logan (10x to 50x), im thinking it has to be Logan. How are they going to achieve that much of a jump otherwise

How exactly are they going to deliver 75* more performance without first turning Tegra into a laptop or desktop computing platform?

They are probably going to do that, Win 8 has already been announced with ARM support.

 

[Exophase]

Just saw the video video.

nVidia claims that there are no consumer devices today that can play 1440p video. And yet I'm sure a high end desktop CPU wouldn't break a sweat doing it in software, much less a GPU. Maybe I'm missing the point of that statement, and that it was really implicitly directed towards SoC hardware. But I don't really get the appeal of playing videos at that resolution in the first place, and I especially don't agree with the laptopmag writer that 2560x1600 is a resolution you'll want on mobile devices. But I don't consider tablets mobile devices, maybe he does. It's a useful resolution in general, but I find > 1080p to be vastly overkill, and it isn't going to matter in 2012 if no one is releasing movies at that bit-rate. It isn't worth the bandwidth or storage space.

I don't see how they got anything up and running for demonstration in 12 days. They said that they were sampling in December, wouldn't that mean they've had at least a couple months?

 

[Ailuros]

Hm, so the GPU is now 12-core now instead of.. what, 8-core for Tegra 2? I never really got what that actually meant. I wish nVidia would talk in more discrete real life terms instead of just made up marketing terms. I wonder what this actually means though. That's only 1.5x the cores nVidia, how much better can it be ;p

It's clear from the Coremark scores that they have no problem pulling numbers straight from the website that were compiled with a much older version of GCC. It's good that they at least mentioned it on the slide, since it's such an obvious gotcha. One interesting thing is that all of the Tegra 2 Coremark submissions get about the same amount, while the Intel ones fluctuate quite a lot. Coremark scores are kind of hard to take seriously all around, but I guess that's not really going to change.

Let's take the KISS approach: whatever the real amount of FP20 PS ALUs in ULP GF/T2, you'd have twice as much on ULP GF/T3 at a higher frequency. VS ALU count most likely stays the same with the amount of TMUs being the next best question mark. Arun may want to shoot me for it, but I still expect a 4:1 ALU:TMU ratio.

 

[MarkoIt]

Kal-El @28nm 5x Tegra 2
Wayne@28nm 2x Tegra 3
Logan@20nm 5x Tegra 4
Stark@20nm 1.5x Tegra 5

We don't have information on Tegra's architecture. It would be cool to have FLOPs number and compare them with desktop graphic cards...

Kal-El could be on the same class of performance of a G86 (8400)?

 

[Ailuros]

Kal-El could be on the same class of performance of a G86 (8400)?

Sorry but that one made me spray my coffee all over the desk. Try a fraction of a ION IGP with less capabilities and you'll be there.

 

[Deleted member 7537]

NGP raw power already beaten?

 

[Deleted member 13524]

Sorry but that one made me spray my coffee all over the desk. Try a fraction of a ION IGP with less capabilities and you'll be there.

The 8400 GS is slower than ION, afaik. Ion uses the refreshed G98 (a bit more performance/sp), not the older G86.

Nonetheless, maybe it's really close to the 8300 GS of G86: 8sp @ 900MHz + 450MHz core.

 

[Brad Grenz]

Those are the nerdiest codenames ever. I love it!

 

[Lazy8s]

Tegra 3 isn't touching NGP in the graphics department. Also, was the 28nm reference a typo in that response above? 40nm, of course.

The next generation will open a gap beyond NGP, certainly.

 

[Ailuros]

The 8400 GS is slower than ION, afaik. Ion uses the refreshed G98 (a bit more performance/sp), not the older G86.

Nonetheless, maybe it's really close to the 8300 GS of G86: 8sp @ 900MHz + 450MHz core.

There are no hot-clocked ALUs in Tegra GPUs (yet) and T3 doesn't sound like it either. When you have a 2:1 ALU frequency:core frequency ratio it's expectable that you get twice the arithmetic throughput.

I myself am confused since I thought T2 has 2 Vec4 PS ALUs but judging from what Anand is repeating over and over again (4 ps + 4 vs and 4 MADDs from each) it's more like 8 FLOPs out of the FP20 PS ALUs * 0.24GHz = 1.92 GFLOPs/s. If those should be 2 Vec4 PS ALUs after all it's 3.84 GFLOPs.

The 8300GS you're mentioning above is at 14.4 GFLOPs and you can multiply the amount of it's TMUs by 450MHz. On Tegra2 you have 2 TMUs * 240MHz = 480MPixels/s fill-rate.

I'd say way too many seem to overestimate embedded GPU designs.

NGP raw power already beaten?

Ok I'll quit drinking even water in front of that PC while browsing these forums

 

[Mike11]

From sample to device in 6 months must be some kind of record, right?

 

[GZ007]

Now thats a realy good joke from nvidia. What is this "performance" what they measure ? Maybe they could show also some linpack numbers for Core 2 duo and tegra 2.

 

[Deleted member 7537]

Ok I'll quit drinking even water in front of that PC while browsing these forums

just asking, lol.

 

[OlegSH]

Now thats a realy good joke from nvidia. What is this "performance" what they measure ? Maybe they could show also some linpack numbers for Core 2 duo and tegra 2.

It should be obvious that they measured T3 FP performance with 4x NEONs versus T2 without it

 

[cjo]

Tegra 3 officially announced; in tablets by August 2011, smartphones by the end of 2011

Project Kal-El:

• World's first mobile quad-core CPU
  

Didn't ARM beat Nvidia to this in the NGP?

 

[AlNom]

Logan must be the successor to Rampage.

edit: Kal-El > *.* That graph is backwards.