Maybe nothing on the market yet, but there are a bunch of announced single A9 products. Off the top of my head, Freescale, ST and Fujitsu all have them
Tegra 3 officially announced; in tablets by August, smartphones by Christmas
- Thread starter Mike11
- Start date
How much slower is a iPhone4 vs. any dual core A9 powered smart-phone for example while browsing?
That NV isn't badly positioned on the CPU side both from a timeframe as well as an application/user experience overall if they come along with a quad A9@2GHz and over half a year earlier in shipping devices than their competitors for the 28nm generation. Their problem will be most likely elsewhere. If TI manages to get OMAP5 or ST Ericsson A9600 in devices at the same time as Tegra4 (not likely IMHO unless TSMC has some very serious problems with 28LP) then of course NV will be in serious trouble. But more because their SoC will lackluster elsewhere than from the CPU side.
Depending on the complexity of the website, quite a bit actually. But what's your point? Speed doesn't matter?
As I said, Krait is out end of this year.
So let me try to understand your argument so far:
1. nVidia may be behind in CPU, but CPU speed doesn't matter.
2. Therefore, nVidia isn't badly positioned on the CPU side.
http://www.anandtech.com/show/4484/htc-droid-incredible-2-review/5
http://www.anandtech.com/show/4484/htc-droid-incredible-2-review/6
Not too shabby for a A8@800MHz vs. 2*A9@1GHz.
Where did I ever state that CPU speed doesn't matter; I clearly said over and over again that it's not the ONLY defining factor or processor within a SoC.
I think you could add a wee bit more wit if you're trying to twist someone's words.
Did you not notice the difference in Browsermark between all the A8's and A9's?
Again, is it your assertion that "CPU speed doesn't matter"?
So how am I supposed to put these two seemingly disconnected statements together:
1. nVidia is not disadvantaged CPU WISE using a quad-A9 compared to a dual-A15. Sure, there's a time-to-market advantage but your exact statement was:
2. Dual A15 will be significantly faster than quad-A9 on the vast vast vast majority of user applications.
Are you disputing #2? Thinking quad-A9 will be able to match a dual-A15 CPU WISE? Or is your first statement not based on nVidia's CPU WISE strengths?
Backpedaling can make you quite defensive, understandable.
We're all agreeing that the combination of nVidia's early introductions of CPU configurations with their high clock rates helps to offset their questionable CPU agenda (not equipping industry-standard NEON, pushing for quad core, disproportionate balance to CPU), so this discussion has lost the plot at this point.
What will be interesting to see is how many phones adopt the quad core variants and what kind of media reception they get with regards to performance and battery life versus the competition at various points in their lifecycle.
What will be interesting to see is how many phones adopt the quad core variants and what kind of media reception they get with regards to performance and battery life versus the competition at various points in their lifecycle.
Yes I did. Still nothing "huge" considering the A4 CPU is clocked at merely 800MHz.
No and I never claimed anything like that.
I put everything and even time to market into a bucket and claimed that there won't be a noticable disadvantage. I said it more than once that if dual A15 SoCs with their projected GPUs and what not appear on shelves at the exact same time as NV's T4, NV will be in serious trouble. That's not ignoring CPU strength but rather acknowledging that's it not the only thing that matters on a SoC.
There are recent announcements probably indirectly driven from NV itself claiming that they'll have the only quad A9 on the market. Not true and as an immediate reaction newsblurbs from TI appear that claim that dual A15 will be by a lot faster than quad A9. TI is neglecting to mention in that case though that OMAP5 will also most likely have an at least twice as strong GPU as Tegra3 which incidentially can be also used in quite a few cases for general purpose tasks. If NV stays with FP20 PS ALUs the latter sounds like a no no until probably the 20nm generation.
If TI won't work for you as an example pick ST Ericcson's Nova A9600 with dual A15's at up to 2.5GHz and a GPU that equals roughly the XBox360 in GPU performance.
It'll never end obviously. It takes two to tango and in that regard you're just as "innocenct" as I am.
Look at it compared to the 1GHz A8's, since that's the same software stack (Android). If you don't consider a ~25% speedup "huge" then I don't know what to say.
I noticed you left out your own quote:
"If hypothetically both square out at around 2.0GHz, I don't see the quad A9 falling that much short if at all."
TI won't be sitting idle. Dual-core 1.8GHz A9's with SGX544 will be out in between T3 and OMAP5. IMO, this is actually a better solution from a CPU performance perspective than quad-A9 at 1.5GHz. Of course, how well the ULP Geforce compared to SGX544 will be a factor as well.
And again, Krait will be released in that timeframe as well.
So nVidia hardly has a gigantic market lead over their competitors. If T3 manages to be in products by August, there will be a good 4-6 month gap, sure. But in the smartphone land, T3 in smartphones would arrive maybe 1-3 months if even before MSM8960 and OMAP4470.
I'm actually wondering how Apple achieved from 4.1 to iOS4.3 such a performance increase with a quite humble A8@800MHz. What it tells me is that Apple has probably a finer balance between its hw and sw and not that a A8@1GHz would break even with a dual A9@1GHz at web browsing. Then come in a whole damn lot of other factors like caches, bandwidth and what not between different SoCs.
A further lack of optimization example is this early showcase of the Blackberry Playbook vs the original iPad in web browsing:
http://www.anandtech.com/show/4266/blackberry-playbook-review/7
I refuse to explain over and over again what I had in the back of my mind.
I'd expect quad A9@2GHz (estimated) in Tegra4 to be honest. T3 is supposed to be up to 5x times faster than T2 and T4 10x times faster than T2 according to NV's own roadmap. Else T4 = 2x T3. Not a lot IMO.
The supposed "12 core" ULP GF in Tegra3 (and if Anand is right that it's still not a unified design) sounds compared to the "8 core" ULP GF in Tegra2 (IMO 1 Vec4 FP20 PS ALU, 1 Vec4 FP32 ALU, 2 TMUs, 8z/stencil 16bit Z precision) something in the neighbourhood of 2 Vec4 PS ALUs, 1 Vec4 FP32 ALU. T2's ULP GF is clocked up to 333MHz in tablets and 300MHz in smart-phones as it seems. If the T3 ULP GF should now be clocked in the 400-450MHz region it doesn't sound like a major increase to me and I'm still wondering where the 5x times overall increase compared to T2 comes from; I recall Arun claiming that they'll justify that one later on.
The good thing with NV's T2 GPU is that it must have quite well tuned drivers. The GPU block looks more than humble both in terms of unit count and capabilities compared to the competition yet it's still not a slouch. This might be a point where their competitors might want to take a deeper look at with their own designs.
However I expect the SGX544 to be somewhere below the SGX543MP2 performance (yet not too far away) otherwise TI wouldn't claim a 2.5x times GPU performance increase compared to the SGX540@307MHz in the 4430.
If I look at the Motorola Droid3 vs. the Motorola Atrix 4G the SGX540@307MHz vs. the ULP GeForce@300MHz (I assume) ends up roughly 35% faster than the latter in GL Benchmark2.0. In 4460 the 540 is clocked at 384MHz according to public documents and the 544 in 4470 adds another strong premium over the last one.
If NV's performance increases for the T4@28nm are accurate then a hypothetical quad A9 at estimated 2.0GHz is the smallest thing to worry about, since OMAP5 goes SGX544MPx.
I'm talking all this time about T4@28nm; I'd consider quad A9 in T3 to be "set" at up to 1.5GHz and not more.
T3 smart-phones and I suppose also future SoCs have the option for smart-phones (AP30?) to be either dual or quad A9. I'd dare to speculate that the majority will opt for the first for smart-phones.
What could be a problem for NV and T4@28nm could be TSMC. I don't even recall if it's on LP or HP, but hopefully TSMC won't have the capacity/yield problems of the past. If everything should go according to their plans even a 4-6 months gap and NV's aggressive marketing being given it's definitely not bad.
Back to T3: I think I've read somewhere that NV expects to sell as many T3's in one quarter than T2 up to that stage. Hopefully I didn't catch that one wrong, but if that's real then I'm of course prepared for some serious kneeslapping.
There's a lot of room to wiggle with an interpreted language. Speedups of 5x are not uncommon (for instance, Android 2.1 -> 2.2). But we were examining the capabilities of the underlying hardware assuming a similar software stack -- albeit different optimizations will affect different architectures differently.
A simple "that's not what I mean" would suffice. The overly defensive tone is simply off-putting.
Exactly how much faith do you put in these "something-x" numbers?
One thing to note is whether that "5x" is applied solely to the graphics or whether nVidia is attempting to apply some overly simple quad CPU = 2x dual CPU formula.
Looking at the Droid 3 benchmarks with the humble SGX540 (though clocked very high), I'd say SGX is improving their layer drivers at a respectable pace. I wouldn't expect improvements from QCOM until the Adreno 300 series. I'm not sure what the state of ARM's Mali driver stack is.
We'll see. Even the 540 right now isn't doing too badly compared to the competition.
If we're talking about T4, then quad-A9 will simply not be competitive. Even at 2.0GHz, it will just barely compare (and in many cases lag behind) 1.5GHz A15 and Kraits. Nothing except really simplistic synthetic benchmarks actually scale to 4 threads perfectly.
Did nVidia say there'd be a dual-core variant of T3? It would make sense but that's quite an ambitious design variant for such an aggressive release date. And if so, then T3 for smartphones will likely lag significantly behind dual-1.8 OMAPs and dual Kraits.
If it's 28LP, there should be no problem in getting yield by the timeframe of T4. 28nm Snapdragons will be shipping in volume well before then. If it's 28HP (and I don't see why it would be at 2.0GHz), then there's more of a risk, but it's not something nVidia's competitors don't have to worry about.
I would guess that it largely assumes how successful Ice Cream Sandwich tablets are. Honeycomb was fairly lackluster and we'll see if ICS can truly make a dent in the tablet market.
Typical marketing. Considering NV's marketing is quite aggressive in general their competitors might learn a "trick" or two.
It's for the entire SoC IMHO, or else creative marketing.
Considering that the 540 consists of 4 Vec2 USC ALUs, 2 TMUs, 8 z/stencil I'd expect it fair even better compared to the ULP GF at the same hypothetical frequency.
If you think that the Droid 3 is at 960*540 and look at its results compared to the iPad2 (despite it being at 1024*768) it seems like the latter could use some driver optimisations too (especially while comparing low level results).
IF they manage to have devices several months on shelves before the others it doesn't sound like such a big problem to me. If I'm right with my gut feeling that there's still a good portion of additional performance lurking in the IMG drivers especially for MPs then what should worry NV in something like the OMAP5 or similar is the SGX544MPx before anything else.
I saw it in a leaked NV slide in the past but don't recall where. http://www.beyond3d.com/content/articles/112/....nope nothing relevant in there. http://www.anandtech.com/show/4181/...-a9s-coming-to-smartphonestablets-this-year/1 No luck here either, but both have several interesting bits about T3 and others.
NV might have had higher expectations for T2 probably because it being the reference design for Honeycomb amongst other things. I'm not in the absolute clear yet about Ice Cream Sandwich but it sounds so far like Google is going to use several reference designs this time.
I don't think any Android devices save for the Nexus Ss and Vibrant are including most of IMG's latest driver improvements. The Droid 3 and Optimus 3D would both score quite a bit higher just with all of the rest of the improvements from those Hummingbird phones.
As the comparisons between iOS and Android in GLBenchmark from the previous generations showed, those scores aren't directly comparable by a long shot.
As the comparisons between iOS and Android in GLBenchmark from the previous generations showed, those scores aren't directly comparable by a long shot.
What is the schedule for T4? It can't be before 2H of 2012. At which point, if it's quad-A9 (regardless of frequency), it really won't be competitive against dual-A15 rocking devices.
I"m confused with the process they're using. For T2 they used 40G and went into production in the same quarter as GF100 from what I recall. If they're using anything like 28LP or HLP then mass production could start earlier than Q1 12'; if it's HP though they might be forced to wait until TSMC's capacities and yields improve.
IIRC Arun had posted in one of the threads that it was on 28 LPG. Im not even sure if thats a real process but i'd imagine they'd be on a low power rather than a high performance process (along with all the other SoC's)
I dont know about your timeframe for mass production though. Its currently slated to sample in December 2011. Even if that happens, the earliest they can hope for mass production is late Q1 2012, probably only Q2
Last edited by a moderator:
This one sounds like a likelier candidate: http://semiaccurate.com/2011/07/19/southern-islands-kepler-and-apples-a6-process-puzzle-outed/
If Charlie is right and AMD manages to manufacture something as complex as SI on HPL or whatever its called in Q4 11' I don't see why a SoC would have a problem around that timeframe or slightly later. Presupposition of course being that the SoC was laid out from the beginning for the specific process.
Could be but i just searched for the post from Arun and i found it - http://forum.beyond3d.com/showpost.php?p=1515753&postcount=452
Now im guessing he meant T4 as we know T3 is on 40 LPG as well.
There should be no problem for production of an SoC in general. For example Qualcomm's MSM 8960 is supposed to be in production in Q4 11' (though i think they're using 28 LP). But if T4 is only going to sample in Dec 2011 then mass production will start in Q1 2012 at the earliest.
Oh gosh, that is embarrassing. My info that it was 28LPG was reliable, but it was also pretty damn old - I should have been a lot more careful about saying something like that, things can change, or maybe even very good sources can be wrong... I think it's fair to say I wasn't the only one that was surprised to learn T3 was still on 40nm.
Mike Rayfield told me explicitly at MWC11 that T4 was on 28nm High-K and he downplayed the 28nm SiON (28LP/LPT) process. I don't know if it's 28HPL or 28HPM - given that I don't expect chips to come back before Q4 2011 they are both possible. I expect 28HPM would be best and TSMC insists there's a lot of interest in it, but if it's 28HPL for time-to-market reasons they might still switch to 28HPM for T5. We'll see. Either way, definitely not 28LPT...
Also when I asked about where the 5x performance claims come from when it used to be 2.5x for T3 versus T2 back in early 2010, he said there's something specific it refers to that they might explain later, and that the specifications for T3 changed since early 2010 anyway. So that makes me think T3 was indeed perhaps on 28LPT as late as one year ago, or maybe I'm just wrong...
Mike Rayfield told me explicitly at MWC11 that T4 was on 28nm High-K and he downplayed the 28nm SiON (28LP/LPT) process. I don't know if it's 28HPL or 28HPM - given that I don't expect chips to come back before Q4 2011 they are both possible. I expect 28HPM would be best and TSMC insists there's a lot of interest in it, but if it's 28HPL for time-to-market reasons they might still switch to 28HPM for T5. We'll see. Either way, definitely not 28LPT...
Also when I asked about where the 5x performance claims come from when it used to be 2.5x for T3 versus T2 back in early 2010, he said there's something specific it refers to that they might explain later, and that the specifications for T3 changed since early 2010 anyway. So that makes me think T3 was indeed perhaps on 28LPT as late as one year ago, or maybe I'm just wrong...
