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

[Ailuros]

Not with Zacate. That one has its GPU clocked at 500MHz.
But Ontario (the 4-9W version) has its "Cedar" clocked at 280MHz, so it should actually be comparable-ish.

Point taken.

Nonetheless, the NGP release should be just a couple of months away from the release of the next-gen low-power APUs on 28nm.

I said that it isn't shipping yet.

In other words, if the NGP ever gets to "compete" with AMD's low power offerings in pure performance, it won't be for long.

It won't be for long even with the embedded SoC 28nm generation.

Cortex A15 + Rogue solutions should be a whole other story, though.
2012 is the year when x86 and ARM solutions will really clash together in apples-to-apples comparisons.

Unless AMD plans to enter the embedded market in 2012 already it might very well happen. As for Intel on the other side we've merely seen bold claims of a self-proclaimed victory.

In any case you don't absolutely need Rogue to beat the NGP GPU block. A SGX54x MP4@400MHz is already good enough for that one.

 

[loekf]

Kel-El demo at Computex:

http://blogs.nvidia.com/2011/05/project-kal-el-demo-previews-future-of-mobile-gaming/?sf1539138=1

Nice ;-)

When you switch to 2 Cortex-A9 cores (from the total of four) the fps drop to 10.

 

[liolio]

Kel-El demo at Computex:

http://blogs.nvidia.com/2011/05/project-kal-el-demo-previews-future-of-mobile-gaming/?sf1539138=1

Nice ;-)

When you switch to 2 Cortex-A9 cores (from the total of four) the fps drop to 10.

You mean that Nvidia optimized PhysicX for the CPU? / Sarcams -----------> []
Look pretty nice

 

[loekf]

You mean that Nvidia optimized PhysicX for the CPU? / Sarcams -----------> []
Look pretty nice

I'm not fully sure what the Tegra GPU is. I picked up once that the GPU in the Tegra 1 & 2 was quite simple, nothing more than a rasterizer, no fancy shaders. Using four Cortex-A9s to play with, you have of course quite a lot of GOPs using NEON instructions for physics (16-way SIMD... ?).

 

[metafor]

That's quite some heavy use of SIMD for the CPU cores to affect fps that much.

 

[Exophase]

I'm not fully sure what the Tegra GPU is. I picked up once that the GPU in the Tegra 1 & 2 was quite simple, nothing more than a rasterizer, no fancy shaders. Using four Cortex-A9s to play with, you have of course quite a lot of GOPs using NEON instructions for physics (16-way SIMD... ?).

Tegra 1/2's GPU is OpenGL ES 2.0 compliant, it has the standard vertex/fragment shading for this hardware class with some extensions.

That of course says nothing for GPU physics; Kal-El is probably an update to the same architecture, with no unified shaders and only FP20 fragment shading (I think, not FP24, will have to check that again later). So naturally it'd be done on the CPU (as is indicated by the article), and apparently that makes this demo CPU limited with two cores active. Note that floating point operations on Cortex-A9's NEON are only two way, not four way.

 

[Deleted member 13524]

Apart from the physics, I found the demo a bit underwhelming, to be honest.

I find the Electopia demo a lot nicer, with all the DoF and bloom post-processing effects. Looks more "up-to-date" to me, at least.

Plus, I find it a bit weird that the framerate seems to be cut in more than half when they halve the available cores. Those 4 cores weren't hitting 100% usage, even when the framerate drops with all the cores activated (~2:50).

 

[KKRT]

Plus, I find it a bit weird that the framerate seems to be cut in more than half when they halve the available cores. Those 4 cores weren't hitting 100% usage, even when the framerate drops with all the cores activated (~2:50).

Probably it was GPU fault.

 

[Erinyes]

Octa Core Wayne? Has Theo actually found the scoop of the day (week?) or is it just the usual - http://www.brightsideofnews.com/news/2011/6/30/nvidias-28nm-wayne-tegra-to-come-in-two-versions.aspx

As per the article, Quad core Wayne(sorry..Robin ) should be sampling by Dec 2011 and will be out in time for the Windows 8 launch. Presumably Cortex A9, and 24 "core" GPU (up from the 12 "core" Kal El)

 

[Ailuros]

Octa Core Wayne? Has Theo actually found the scoop of the day (week?) or is it just the usual - http://www.brightsideofnews.com/news/2011/6/30/nvidias-28nm-wayne-tegra-to-come-in-two-versions.aspx

As per the article, Quad core Wayne(sorry..Robin ) should be sampling by Dec 2011 and will be out in time for the Windows 8 launch. Presumably Cortex A9, and 24 "core" GPU (up from the 12 "core" Kal El)

 

[metafor]

Octa Core Wayne? Has Theo actually found the scoop of the day (week?) or is it just the usual - http://www.brightsideofnews.com/news/2011/6/30/nvidias-28nm-wayne-tegra-to-come-in-two-versions.aspx

As per the article, Quad core Wayne(sorry..Robin ) should be sampling by Dec 2011 and will be out in time for the Windows 8 launch. Presumably Cortex A9, and 24 "core" GPU (up from the 12 "core" Kal El)

Wouldn't a 24-core Geforce mobile be just enough to catch up to the SGX543MP2? With potential MP4 implementations coming for tablets (not to mention AMD's Fusion and Intel's next-gen GM), would that really be competitive?

I don't know how well it'd scale to smartphones, even with the shrink to 28nm.

 

[Ailuros]

Wouldn't a 24-core Geforce mobile be just enough to catch up to the SGX543MP2?

Let's stop counting vector lanes as "cores" please. If we'd have to get into that kind of silly season the MP2 would count more "cores" and more efficient ones.

With potential MP4 implementations coming for tablets (not to mention AMD's Fusion and Intel's next-gen GM), would that really be competitive?

If that BSN newsblurb is even an inch close to reality (which I severely doubt) then NV's roadmap is most likely screwed up. Tegra3 is according to that supposed to be 5x times faster as a SoC compared to Tegra2 and Tegra4 (Wayne) 10x times Tegra2. If you divide 10 with 5 you get 2x yes? Meaning NV projects Tegra4 to be only 2x times faster than Tegra3.

From all the so far leaked infos T3 looks like a quad core A9 CPU@1.5GHz with a ULP GF clocked at =/>400MHz consisting this time of 2 Vec4 PS ALUs instead of only 1 in Tegra2.

If NV is not bullshitting on purpose with that road-map I can easily imagine what Tegra4 would contain.

Now BSN comes along and claims an 8 core A15 with twice as many ALU lanes than in T3 and the entire enchilada is only be 2x times faster? Why doesn't BSN reveal the frequencies of that miraculous constellation; let's see a couple of hundred MHz only for the equation to come to the correct result and we'll have even more reason to laugh about it.

Your over-estimating both Tegra4 as well as other SoCs as it seems. Don't forget that unless NV's competitors speed up their road-maps NVIDIA has good chances to a LOT earlier than the others with its 28nm SoC.

I don't know how well it'd scale to smartphones, even with the shrink to 28nm.

BSN stands for bright side of news yes? Alas if you'd read the BS in another way

 

[ninelven]

Now BSN comes along and claims an 8 core A15 with twice as many ALU lanes than in T3 and the entire enchilada is only be 2x times faster? Why doesn't BSN reveal the frequencies of that miraculous constellation; let's see a couple of hundred MHz only for the equation to come to the correct result and we'll have even more reason to laugh about it.

Did you even read the article? It states there is to be 2 versions of Wayne. Clearly the "Robin" version with 4 A15 cores @ ~1.5Ghz is the one meant to be 2x faster than Kal-El...I'm not saying they are correct, but if you are going to rake them over the coals, the least you could do is actually read the article carefully.

 

[metafor]

Let's stop counting vector lanes as "cores" please. If we'd have to get into that kind of silly season the MP2 would count more "cores" and more efficient ones.

nVidia deems it sufficient to dub them "cores", so I don't see why not. In either case, my point was that if you double the number of shaders in Kal-el, that still would be enough just to catch up to SGX543MP2.

 

[Exophase]

Did you even read the article? It states there is to be 2 versions of Wayne. Clearly the "Robin" version with 4 A15 cores @ ~1.5Ghz is the one meant to be 2x faster than Kal-El...I'm not saying they are correct, but if you are going to rake them over the coals, the least you could do is actually read the article carefully.

The article talks about a higher end 8 core version of the same chip for tablets. An 8-core Cortex-A15 is an outrageous prospect and I doubt nVidia has the die budget for it at 28nm (Kal-El is 40nm, so any alternate for Wayne would 28nm). An 8-core A9 is feasible but would be a very poor design decision. Really I have a hard time imagining 8-core CPU anything providing a worthwhile performance advantage in most usage scenarios in 2013 tablets. In fact, I think even quad core will be highly underutilized in late 2011/2012 tablets. Hence why TI and ST-Ericsson will be offering dual core A15s.

I'd also like to know how in the world a 1.5GHz quad-core A15 will offer triple digit GFLOPs.. Most likely A15 NEON will do 4x FMA per cycle, with 8x being an unlikely possibility.. even then we're only looking at 1.5 * 8 * 2 * 4 = 96 GFLOPs. But 48 is much more realistic. Or is this supposed to include GPU too? Then just how high can those 24 ALUs (48 FLOP/cycle presumably) clock?

 

[ninelven]

The article talks about a higher end 8 core version of the same chip for tablets.

And notebooks... die size isn't really an issue (physically) as long as it is a product worth paying for...

 

[metafor]

The article talks about a higher end 8 core version of the same chip for tablets. An 8-core Cortex-A15 is an outrageous prospect and I doubt nVidia has the die budget for it at 28nm (Kal-El is 40nm, so any alternate for Wayne would 28nm). An 8-core A9 is feasible but would be a very poor design decision. Really I have a hard time imagining 8-core CPU anything providing a worthwhile performance advantage in most usage scenarios in 2013 tablets. In fact, I think even quad core will be highly underutilized in late 2011/2012 tablets. Hence why TI and ST-Ericsson will be offering dual core A15s.

nVidia has never been one to shy away from having designing something to sound good with marketing. Most laymen hear quad-core and automatically expect it to be double the performance of any dual-core; regardless of the type of CPU.

I mean, really, was the ultra-high-def video handling capabilities of Kal-el really necessary in any usage scenario?

 

[Exophase]

And notebooks... die size isn't really an issue (physically) as long as it is a product worth paying for...

Netbooks (and moreso proper laptops) are an uphill battle for ARM SoCs. They need all the single threaded performance they can get to be viable, and trading that for a ridiculous number of cores is not a good compromise. And unlike Intel mainstream x86 CPUs there's much less opportunity for high end binning to soak up die production costs. Tegra 2 was quite conservative in die size, I don't see nVidia pulling a 180 on that.

nVidia has never been one to shy away from having designing something to sound good with marketing. Most laymen hear quad-core and automatically expect it to be double the performance of any dual-core; regardless of the type of CPU.

I mean, really, was the ultra-high-def video handling capabilities of Kal-el really necessary in any usage scenario?

Design for marketing is quite unfortunate. At least with quad core there'll be some potential benefit for games in the long term, and PSV will help push that along. 8 core is just crazy, one need only look at the progression of core counts on the desktop to see how viable that looks in the tablet world where there's less interest in heavy lifting background parallel tasks and even long-term multitasking in general.

The ultra-high def stuff is just silly. Who is going to even have 1440p or whatever TVs?

 

[metafor]

Netbooks (and moreso proper laptops) are an uphill battle for ARM SoCs. They need all the single threaded performance they can get to be viable, and trading that for a ridiculous number of cores is not a good compromise. And unlike Intel mainstream x86 CPUs there's much less opportunity for high end binning to soak up die production costs. Tegra 2 was quite conservative in die size, I don't see nVidia pulling a 180 on that.

A15 (and Krait at 2.5GHz) should provide quite substantial and competitive single-thread performance; enough for netbooks, I'd say. Of course, we don't know how the next-gen Atom will perform, so ARM offerings could very well still have an uphill battle.

Design for marketing is quite unfortunate. At least with quad core there'll be some potential benefit for games in the long term, and PSV will help push that along. 8 core is just crazy, one need only look at the progression of core counts on the desktop to see how viable that looks in the tablet world where there's less interest in heavy lifting background parallel tasks and even long-term multitasking in general.

I really have a difficult time imagining a use for quad-core in a gaming scenario. I could see 2xA5 + 2xA15 as a compelling quad-core solution as that would allow low-power during non-intensive use-cases. I'm not even convinced modern desktop quad-core solutions are being utilized in any sort of meaningful way outside of heavy workstation workloads such as 3D scene rendering.

On a tablet? It's probably better to utilize a GP-CPU -- which is the direction the mobile GPU's seem to be going -- or even the DSP for media-intensive tasks such as photo-editing.

The ultra-high def stuff is just silly. Who is going to even have 1440p or whatever TVs?

But hey, go on most blog sites and watch how people anticipate the 1440p video capabilities and tout how "fast" Kal-el is.

 

[Lazy8s]

The higher definitions for video is just that push towards "retina" displays in tablets and netbooks.