Nvidia Tegra

Wow. I consider myself probably slightly biased in NVIDIA's favour sometimes, so I'm not really used to having to argue against someone with such... errr... incredible optimism.

Is that something you've heard in a reliable way or are you just mixing real info with speculation?
- I am quite confident that NVIDIA never did any work for Tegra on Symbian, and you're probably the very first person ever to claim that publicly.
- Tegra1 was originally aimed exclusively at WinCE, WM6, and WM7 (aka Photon) as Ike Turner pointed out.
- Even though they were part of the Open Handset Alliance, they did not start porting to Android until late 2008 after WM7 was cancelled completely. They specifically bragged about how little time the port took them at MWC09.
- For Tegra2, they decided to focus on Android and WinCE (see: the original Asus tablet project with Tegra2 and WinCE 7) although they did toy with Ubuntu and Chrome.

What in the world are you talking about? The MSM8260 started sampling in Q2 2010. There must be plenty of prototypes and design-ins - the fact you don't know about them is irrelevant. Whether there are true design wins (i.e. fully locked and guaranteed to come to market) I don't know, but these things don't get decided overnight. OEMs start evaluating these chips long in advance and have tentative phones with them in the pipeline that they cancel if they decide the platform isn't worth it. The question is not whether they have any customer interest; of course they do. The question is how much of that interest will materialise into real products. Just look at Tegra1...

As of a few months ago, I'm pretty sure there were three dual-core competitors with real shipping products expected in Q2 2011: the ST-Ericsson U8500, TI OMAP4, and Qualcomm MSM8260. The MSM8260 is the one which started sampling last, so I agree it's very likely to slip to Q3 or even Q4. But pretending all three are nearly certainly going to slip in H2 is ridiculous.

NVIDIA did not "execute perfectly". They certainly expected products out even earlier and given their software maturity in mid-2010 I don't think it's just their partners who failed to deliver. But they did execute better than their competitors, and that's what matters most in practice.

As for first shipping Tegra3 products in 2H12? That would be a massive delay compared to their earlier targets. I assume you're just speculating randomly here and don't actually know... but we'll see.

Apple has no reason to look at Tegra 2 except for competitive intelligence reasons, so I assume that list is just speculation? The real list would also be ridiculously bigger than that.

You really need to stop considering phones and tablets together. It's obviously very clear that many large OEMs including Samsung will ship Tegra2 tablets. My point is that when it comes to phones, LG and Motorola are the only Tier1s that are likely to release anything (with the possible exception of HP/Palm but I'd rather bet on TI or QCOM there) although we'll obviously see some more Tier2s.

That's the 3G version. BB has been very clear in saying that the WiFi-only version would come first. It's very very very unlikely that doesn't come out before the end of Q2 2011 and possibly quite a bit earlier.

That's by far the most insane statement I've heard so far this year, I hope you don't seriously believe that crap. It's certainly true that Tegra3 will be available before OMAP5, but that doesn't warrant looking at it that way in the slightest. Once again you seem to reason as if design wins were decided just a few months before the product is released - nothing could be further from the truth especially for phones.

Tegra3 on 28nm is already sampling to potential customers AFAIK (just like Tegra2 started sampling in July 2009). There is no need for a Tegra 2.5 launching in Q2 - your timeframes are wrong. What I can easily imagine, however, would be a higher-clocked 1.2GHz Tegra2 SKU (exact same chip or maybe a respin) for tablets. Anything more than that just isn't going to happen.

Well it's not my job, but that doesn't prevent me from having a go at it anyway!

 

There you go again, you are comparing a PHYSICAL PRODUCT hitting the shelves with an APPLICATION PROCESSOR's "details start to filter out", after having just made the point about how "unfair" it was to make such comparisons.

(and thats ignoring your unreasonably pessimistic launch date for playbook)

 

4G, actually.

 

Eee Memo is using the 8260 I think. June.
So make that 3 months?

 

Bah, I don't care what the ITU says. I want my 1Gbps peak speed (aka 100Kbps on a congested network) or I refuse to call it anything else than 3G (more seriously, I think the most realistic definition of 4G is "anything that uses MIMO and/or more than 5MHz channels" in which case WiMax/LTE fit but HSPA+ doesn't before it hits 42Mbps - not that I expect anyone to care, oh well).
---
And good point about the Asus Eee MeMO, I had seen Q2/Q3 somewhere but a quick Google tells me June is definitely their ETA for the first SKU. So it's even more absurd than I thought to claim the MSM8260 doesn't have any design wins...

BTW, a technical point on Tegra: I do believe NV when they say their memory controller is extremely high efficiency given their experience there and the rather subpar performance of some competitors in the LPDDR1 generation, but nonetheless they are competing against 64-bit 800MHz LPDDR2 with only 32-bit 600MHz LPDDR2 (or 667MHz DDR2 on tablets). And the QSD8672 supports up to 64-bit 1333MHz DDR3! I would be very surprised if NVIDIA wasn't often bandwidth limited when pushing both the CPU and GPU at the same time (i.e. gaming)

 

They pretty much have to. An extra memory channel not only inflates power dramatically, it also inflates die size. 8672 will never fit into a smartphone's power envelope or its cost structure. I'm not sure about OMAP4 but I don't think there's a single smartphone design out there with it.

Granted, tablets are a different story but I don't think nVidia wanted to make two iterations of Tegra 2 and instead opted for lower performance on the tablet side to be able to service both form factors.

Was it a smart call? Well....who's out to market first

OMAP4 may be faster and I'd venture to say 8260 may even overtake Tegra 2 in some things but none of that matters if it's 3-6 months late.

 

metafor: Maybe the information I have is wrong, but I have seen a presentation where the MSM7x30 and MSM8x55 (same chip afaict) are both listed as having 14x14 PoP Dual-Channel LPDDR2 packages. And those are much lower-end than the MSM8260, which is simply listed as 14x14 without PoP (?!) - or maybe Qualcomm thinks of x16 as a channel *shrugs* That'd be pretty ridiculous. The ST-Ericsson U8500 is another 32-bit LPDDR2 solution but at least it supports 800MHz.

 

Both 7x30 and 8x55 are dual channel but not the 8260/8660. I guess they gave up the second channel for a second core and ran it at a faster frequency.

 

Arun,

I just removed the tomatoes from my eyes. The diagram states "GPU pipeline" so that's 1 Vec4 PS + 1 Vec4 VS ALU per pipeline, ie 2 Vec4 for each.

 

Why "none of that matters"?

What if, in 6 months, I said: "Tegra3 may be faster than OMAP4, but none of that matters if it's 3-6 months late.", would you agree or disagree?

If two competitors have 12 month schedules, relatively offset by 6 months, each leapfrogging the other with every release, how does that leave TI at a permanent disadvantage and Nvidia at a permanent advantage?

 

It's not a permanent advantage so much as a lot of revenue lost to TI/Qualcomm. And we're not talking leapfrog type of performance advantage here. OMAP4 may be somewhat faster than Tegra 2 in some cases, but the baseline is more or less the same dual A9.

Snapdragon's in an even worse shape due to a closer-to-A8 class processor (though it helps that it scales to higher frequencies).

 

Obscure design choices FTW! Hmm, maybe the 7x30 supports dual channels so that it's not bandwidth starved when using LPDDR1 and they didn't bother trying to save a little bit of money by only supporting LPDDR2 on one of the two channels. 64-bit LPDDR1 would be cheaper than 32-bit LPDDR2 when it comes to the memory chips but take more power - that and capacity constraints is why the Apple A4 uses 64-bit LPDDR1 afaik.

I wonder if the MSM8260 is about the maximum performance level you can reasonably hope for without being quite bandwidth constrained on 800MHz LPDDR2. I am also reminded it only has 512KB L2 versus 1MB L2 for all the 2xA9 SoCs I'm aware of...

Ah, that makes a lot more sense, should teach both of us for reading through these things too fast

 

Ya. Like I said, memory controllers are expensive But like nVidia, in order to fit a dual-core chip in a smartphone type of die (both size-wise and power-wise), compromises had to be made. For power-house chips that are more on-par with the OMAP4 in tablets, people will have to wait for 8960

 

Pretty sure that is 5 months. There is a hands on with it here:

http://www.pcmag.com/article2/0,2817,2375327,00.asp

Quite a few interesting quotes in there. Seems very much an odd duck.

Anything else announced yet?

 

Nothing else QSD/MSM other than the single core MSM8250A.
Weren't all TG2 tablets shipping in Q1/"March" (ahha) due to Honeycomb anyway?

We'll probably see more duallie snapdragon action in MWC in Feb- judging by the article it doesn't seem ready for show, and was there more to showcase the whole lineup. Seems like CES has mostly products targeted for Q1 and MWC for Q2/rest of H2.

 

From the link:

A9...sure...gotta love clueless "journalism....

 

As Samsung showed with the Hummingbird platform, if worn right, the performance crown in the SoC market can net a lot of sales.

The problem for Tegra2 is that it doesn't outclass the upclocked, forthcoming Hummingbird update so much nor will it remain atop when OMAP4 hits, and the A5 will (deservedly) win the perception game, although not a direct competitor.

 

Uhm, I very much doubt that. A single-core 1GHz A9 should usually be slightly faster than a 1.2GHz A8, and as far as I can tell a dual-core A9 will be 30 to 60% faster for web browsing than a single-core A9. If that's not significant, I don't know what is.

Completely agreed from a technical perspective, but *if* NVIDIA and their partners execute, Tegra3 will be available in tablets before the end of the year. So in terms of perception I don't think it's that simple.

 

While CPU benchmarks will definitely show a significant advantage for dual A9s, I don't expect a Gameloft or other high-end 3D Android game nor even a web page display browser speed test to show Galaxy S devices at much of a disadvantage.

 

I agree that there won't be a huge difference 3D-wise (assuming equal quality drivers/very recent software on both and developers not optimising more for one than the other) although I think NVIDIA's marketing (including the Tegra Zone) will make most people think otherwise.

And I disagree about web browsing. I believed for a very long time that dual-core wouldn't really matter for web browsing until everyone started working a lot harder on the problem, but I recently looked into it quite a bit and it turns out a lot of the necessary steps have already happened behind the scenes. It's still far from perfect scaling (it'll improve somewhat over time) but you'd be very wrong to think there's no difference. Don't have the time to get into that here though...