It looks like nVidia is trying to play things a little bit differently with Tegra than it was with all the secrecy surrounding the desktop graphics parts.
Open roadmaps are very important for developers, thus it is no wonder that Intel is considered as the world’s leader: the company is comfortable with publicizing their roadmaps for products years in advance [who didn't know about Hyper-Threading in Nehalem generation back in 2006? 45nm in 2007, 32nm in 2010 - and the list goes on].
nVidia still prosecutes people who disclose something as important and industry changing as changing its architecture from SIMD to MIMD-like, yet it expects that developers will adapt in a matter of seconds – guys’n’girls, grow up. You expect that companies make million-dollar business decisions based on your secrecy – it doesn’t work like that, sorry. Getting back on the subject, the information about the next-generation Tegra parts goes back as far as nVision 08, when we chatted with numerous engineers from newly founded business units such as the "Automotive division", "Telecom business" etc.
The roadmap for Tegra is not covered in a shroud of secrecy with nVidia pushing for a one year life-cycle unlike the current competition which prefers longer lead times. There are two generations in the works, with design papers being written for generations to come. NVidia is trying to have a next generation of Tegra debuting on first day of Mobile World Congress. In case of T2 and T3 [codenames for Tegra Gen2 and Gen3], the parts should debut on February 15th, 2010 and February 14th, 2011.
nVidia’s roadmap discloses a 400% performance increase between Tegra Gen2 and the current generation.
The second generation Tegra part will offer a 400% performance increase in terms of performance. While this "performance increase" is not specified, we know that the company will utilize parts that are designed for TSMC’s low-power process. Thus, you can expect to see a next-gen ARM core [Cortex instead of ARM11] in a dual-core form, and a GPU subsystem based on the GeForce 9 architecture. We expect that nVidia will tout CUDA and OpenCL compliancy with the Tegra Gen2. Don’t even get us started with Tegra Gen3 – the GPU on that will be more than a beast.
Truth to be told, GPU is already utilized much more inside Tegra than it is on the PC side of the fence: image processing, video processing, transcoding, OS acceleration, application acceleration – all this is being utilized inside mobile operating systems – every miliWatt counts here.
When it comes to manufacturing processes, don’t expect earth shattering moves. First and foremost, the parts have to come on a reliable manufacturing process, otherwise OEM and ODM vendors won’t even look at you. According to information we have at hands, we believe that Tegra 2nd Gen will be manufactured using 40nm Low-Power process. At present, sources are refusing to comment will the part be manufactured in TSMC or GlobalFoundries, but given the timeline of the planned launch [February 15th, 2010 - Mobile World Congress 2010 in Barcelona, Spain], we think that TSMC is in the box seat for this one too.
Given the current die size of 144mm2 [Tegra 600 and 650], logic dictates that Tegra’s transistor budget is inside the 150mm2 die space. But if nVidia does not reduce the 150mm2 die size budget and manufactures the part in 40nm LP, Tegra Gen2 could squeeze as much as 850 million transistors, or more than whole Core i7 processor! Bear in mind that ATI packed 826 million transistors in 136mm2 using TSMC’s 40nm process.
Naturally, the comparison stands only if Tegra designs venture in the CPU and GPU-levels of multiple silicon layers. Then again, SoC is a more complex design than a CPU or a GPU – it has to pack both. But with 40nm and 28nm low-power processes, Tegra Gen2 and Gen3 could punch as much as billion transistors in a mobile package. If the engineers are able to pull this off, we could see the birth of completely new computers. After all, it isn’t impossible to imagine a cellphone packing enough computing power to play modern games in handheld resolutions.