Discussion in 'Mobile Devices and SoCs' started by Arun, Feb 22, 2011.
My first post-MWC11 article is finally up - enjoy!
Gee Arun, did you buy a time machine @MWC2011? This speed is positively shocking here @B3D.
Nice piece though. Eagerly waiting for your gpu article.
Nice round up of key players for the next generation.
Is Broadcom considering licensing new GPU IP? They seemed like they might be trying to push ahead with more of their own stuff for future VideoCore tech.
One clarification on the article: August is obviously for WiFi-only Kal-El tablets. 3G tablets are not strictly impossible by then (some operators have extremely fast certification processes with module-based devices) but either way it's not really what you should be looking for.
No, nothing so fanciful - I just set a very aggressive deadline, that way I'd still be early even if I was late (which I was) As for the GPU article, I'm hoping to start with a more in-depth article on at least one GPU architecture. We'll see if that happens or not (it's not just up to me) and whether I can eventually get sufficient information on some other architectures to do something similar for them (I think it's worth a delay personally). I can always stick to the previously expected overview otherwise.
I think their current GPU IP based on the AlphaMosaic Vec16 architecture and accelerators is surprisingly fast and solid for something that's been developed in-house by a non-GPU company and with a very unusual architecture (Larrabee should be jealous, not as software-centric but still).
I guess I sounded maybe a little more harsh than I intended in the article. The BCM21654 seems to use a single-core variant for both VGA video and 3D (BCM2727 is dual-core, BCM2763 is quad-core) and if it's cost efficient that makes a lot of sense for them. In theory they start with an advantage versus all other 3D architectures because it can reuse some of that silicon for video/imaging - but who knows how efficient the overall architecture is in practice.
I'm not convinced making their own GPU IP is worth the investment level and risk even in the low-end, but given that they've managed to come up with something decent performance-wise they might as well continue to do so for now. On the other hand, they've also licensed the Cortex-A15, and I think they would be at a big disadvantage if they tried to stick to their in-house architecture (instead of Rogue or T604 or ...) for the high-end. So I'd expect them to be considering 3rd Party GPU IP going forward but probably not from top-to-bottom, at least not for some time.
Tegra 2 is only $15?
On average probably even less. From one side the higher the order quantity from potential buyer X the more will X push for lower prices and from the other side what do you think TI is selling its OMAP4430 exactly for? Minimum quantity for a Tegra2 order if memory serves well is 100.000.
Apps processors don't really cost that much. I think the iPhone4 BOM showed the A4 as ~$8. This is the market Intel wants to compete in.
Speaking of Intel, did they show anything apart from that lovely quote?
Arun, do you know how many cores?
$15 for high volume smartphone OEMs (i.e. a few million units) but closer to $30 for low volume tablets (i.e. 100K?) AFAIK - given the higher die size for Kal-El, we might be closer to $20 for smartphones I guess.
As for the A4's cost, $8 is probably realistic because it is Apple's own SoC and they don't have an intermediary taking 40-50% gross margins. If it wasn't for that, I'd expect it to only be very slightly cheaper than Tegra 2 (53mm² on 45nm with 2x32-bit memory bus isn't free).
I had a good discussion with someone from the ex-Infineon business, don't think that's what you had in mind They showed MeeGo tablet stuff a bit, but Medfield was behind closed doors only, and they didn't reveal much of anything at the press conference which I didn't attend. I'm a bit less optimistic about their prospects than I was one month ago (given what happened at Nokia and the lack of anything interesting from Intel at MWC), although the good news for them is TSMC 28nm is a slightly behind where I thought it would be by now.
I suspect the number of TMUs and ALUs is probably a more important factor than the number of cores, don't you think?
For TMUs, ALUs even numbers make only sense; while core amount can be either an even or uneven number/amount. The most meaningless contribution of the day and no I doubt anyone knows yet apart from insiders.
Try running with a big chunk of bacon in front of Rys; it might work
TI recently stated during an investor conference that the overall OMAP ASP is around $15 with OMAP4 closer to, or a little above, $20.
The geforce ULV pixel shaders are actualy pixel/blend shaders. They are also ROP-s. So if they double the pixel/blend shaders they end up also with double fillrate.
Did you see any tablet/slate amd products ? I'm really interested in a bobcat tablet towards the end of the year.
It is my understanding that there's one programmable blending unit in the pixel shader block, which if the core isn't blending can be used for an additional single FLOP. T2 ULP GeForce is also capable of 8 z/stencil, which if it should double it brings it on par with a single core SGX543 (16z/stencil); and that's only double the z/stencil fill-rate and not texel fill-rate.