Nvidia Tegra 3 started out pretty nicely early this year, but it was quickly beaten by chips like Qualcomm S4, Exynos 4 Quad and Apple's A5X, and then it was barely even a top 3 chip in the Android world. This is why Nvidia decided early on that they would reposition their chip as a high-end chip for mid-end devices, and that's how they managed to get pretty good marketing for the devices that cost like $200 but had the quad core Tegra 3 in them.
But why couldn't Nvidia, with so much experience in making GPU's, be able to be at the top of the market with Tegra 3? It seems Apple and the others caught them off-guard, because they didn't think they would make chips this fast, this soon. So instead of making bigger chips, and also on 28nm, they decided to take it easy and make a significantly smaller 80mm2 chip (vs 120mm2 the others), and also on 40nm, instead of 28nm, which means that they effectively built a 40mm2 chip at 28nm, when all the others had 120mm2 chips.
The CPU's were fine, but it was in GPU's, their area of expertise, where they got beaten badly, because they made so few cores for it – Nvidia cores that is, because the others have fewer cores but are more powerful. But GPU cores are very scalable, so they could've still competed if they made more of them, but they didn't.
It looks like things will be a lot different with Tegra 4, and they are going all-in. Initially, Tegra 3 was supposed to be a 2011 chip, and then we'd get Tegra 4 in 2012. But with each generation, Nvidia fell behind a few months, and Tegra 3 ended up being a 2012 chip, and with a huge window until the next Tegra arrives – which is what they call Tegra 4 now, but was supposed to be Tegra 5.
Tegra 4 is supposed to be a completely new architecture, in both CPU and GPU. It will have a quad core Cortex A15 CPU, along with another companion core, probably still Cortex A9, and not Cortex A7. I consider this a pretty big downside when it comes to battery life, and compared to the quad core Cortex A7 cluster Samsung is supposed to have next year in their chips, but perhaps Cortex A7 wasn't arriving fast enough for Tegra 4, so Nvidia decided to skip it. Another explanation is that they already have their own switching technology reliant on a Cortex A9 core, and they decided against using ARM's big.Little.
But the new leaks show us the most exciting part about this new chip – that it will have 72 GPU cores, which is 6x as many as Tegra 3 has, which means it should have 6x or higher the performance of Tegra 3, as GPU cores are very scalable like that. This is exactly the kind of performance jump Nvidia needs in its GPU's for next year, if they want to be competitive, because this year Tegra 3 has been left way behind, and they need to compete not with this year's chips, but with the ones appearing next year, which will be even more powerful.
Tegra 4 will also be able to decode 1440p video in both VP8 and h.264 formats. It will support 2560×1600 resolutions (and even 4k, but I wouldn't bet on running perfect with that), as well as 1080p at 120 Hz. The chip will also support USB 3.0, which is something Exynos 5 supports as well, and I'm sure all high-end chips next year. The chip will finally have fast dual-channel memory, that can be DDR3L or LPDDR3.
I expect the GPU to also have support for OpenGL ES 3.0, OpenCL and probably even CUDA. This is most likely a new GPU architecture from Nvidia that also has unified shaders, so it should be inline with Adreno 300 and Mali T600 series. The timing of the release of this chip will get to determine whether this chip is ahead of its time and better than everything else on the market, or merely competitive with whatever is coming out next year.