It's definitely Tensilica Xtensa-based but that doesn't mean much since that's a highly configurable core. Snapdragon1 is probably also using XTensa for 720p (or maybe not and it was really only Freescale or STM who had licensed that last generation) but its configuration must surely be quite different to achieve 1080p Encode and the much higher bitrates. You could argue it's still incremental, and that'd be fair, but my point is I fail to see how Snapdragon3 could be a more fundamental change. I'm not sure whether it supports CABAC/High Profile (I think it does) but beyond that there's not much they're likely to add besides dual-stream support, which basically just requires copy-pasting hardware or doubling the clock speed (and adding some more dedicated RAM if it's the latter).New video codecs, performance improvements? I thought that it was still the same tensilica xtensa core...
Yeah, it's z430 vs z460. One key factor that hasn't been discussed much is that while z430 was basically limited to one pipeline, z460 is more flexible and (before being acquired by Qualcomm) a company that'd have really insisted for a different pipeline count (or even a different ALU:TEX ratio or whatever) could have gotten it. It's not quite like SGX where any given IP is a very optimized but inflexible solution; here I've heard from two different people that their RTL is apparently flexible enough to change these things as required even without a Mali400 or SGX543-like core concept.Isn't the GPU based on quad-pipeline z430? Snapdragon2 is supposed to be 4x faster than the snapdragon1 so it would seem so.
I don't know about Snapdragon2's fillrate exactly (triangle performance is massively higher though) but the most likely option remains a 4-pipeline solution at slightly higher clock speeds than Snapdragon1. For Snapdragon3, I have no idea but I'd assume a newer revision of z460 (or something like it) with 8 pipelines would make sense. Or maybe not. Sorry, not very useful I know
Not really as I implied above - just there are clear architectural changes in these blocks from SD1 to SD2, whereas SD3 *might* be simply stronger through brute force. Other important changes in Snapdragon2 include things like a probably implemented/used power island for audio decode (ala Tegra and OMAP4) and apparently better clock speed optimizations for the CPU core. It's not surprising they could improve it after their first try, but remember it's not process related because there's little to no performance boost from 65LP to 45/40LP!Are my assumptions wrong?
One other thing to keep in mind for both Tegra3 and Snapdragon3 is that 28LPG is a triple gate oxide process, which means they can use high performance/high leakage transistors *selectively* where they want to in order to achieve higher clock speeds everywhere and much higher clock speeds in places where they're mostly OK with higher leakage like the CPU core. If this is used properly, maybe we'll be pleasantly surprised by the clock speeds.
How are they so wildly overly optimistic? TSMC wouldn't claim the first tape-outs will happen in Q1 (and more in Q2) 2010 if their customers weren't ready for it. This is not the 28HP process, where TSMC wouldn't be ready anyway. Remember: the first 45LP tape-out happened by Qualcomm in mid-2007. This is 2.5-3 years later! Handheld tape-outs always happen early and take a long time to come out in real products. OMAP3 taped-out in August 2006; it took nearly 3 years before the first phone based on it came out!These Semis continue to be wildly, overly optimistic on their timetable for process transition, now for the 28nm generation.