NVIDIA Tegra Architecture

[Silent_Buddha]

Right, because the scale doesn't start at 0.5 or 0.8 like the AMD marketing graphs?

In Sunspider, Google Octane, Kraken, and Vellamo Metal, Anandtech has already confirmed that quad-core S600 is ~ 30% faster than quad-core S4 Pro. Quad-core Cortex A15 will be way faster than quad-core Krait 300 in a wide variety of different applications.

What was the point in bringing AMD into this?

You could have just said just like the Nvidia marketing graphs? As they do and have done the same thing for the past decade. Both companies are doing it.

I'm just confused as to why AMD even needed to be mentioned here. At least mentioning Nvidia would still fit within the subject of this thread, but AMD doesn't have a competitor to Tegra, nor do they have anything to do with it.

Regards,
SB

 

[Exophase]

The smaller die size of the T4 GPU relative to it's competition is certainly a combination of both stripped out forward-looking features and higher GPU clock operating frequency. But the GPU clock operating frequency is not out of this world either. For instance, T4 has a GPU clock operating frequency that is "only" 26% higher than the 5XT GPU used in the SGS4 international variant.

There are lots of things you can do to favor perf/mm^2 over perf/W. Let's go back to tiling for a minute.. the tile memory takes up a substantial amount of extra space. As does the tiling engine itself. And I'm going to guess that all around it's a fair bit more complex to implement (this is totally orthogonal to early Z, most use early Z and tiling). Yet this is what four out of the five major players in the mobile GPU space have chosen. There has got to be a reason for this.

And "only" 26% is actually kind of a lot.

When I look at Tegra 4's GPU I see some very strange balancing choices. Compared to Tegra 3 they've increased both fragment and vertex processing capability by 8x, yet have only increased throughput of the triangle engine (setup, clipping, etc) by 1.5x. I can't begin to explain the rationale behind this one.

That may be true, but we have yet to see any actual perf/w data on quad-core Krait and quad-core Cortex A15 at 28nm, and there are some applications where the performance of Cortex A15 is way ahead of Krait 300 by a factor of 2x or more: http://www.hardwareluxx.de/images/stories/newsbilder/aschilling/2013/mwc/tegra4-press-3.jpg

No point pretending that core perf/W changes with the number of cores you put on it. And while nVidia might have much more efficient Cortex-A15s on TSMC 28nm than Samsung's 32nm ones I kind of doubt it.

Now that slide, that's just grade A bullshit. Just compare the number nVidia gives for Sunspider on Tegra 4, 499ms, with what a 1.9GHz S600 gets: http://www.gsmarena.com/samsung_galaxy_s4_vs_htc_one-review-913p6.php

Difference is only 1.62x, so where does 2.5x come from, aside from nVidia's imagination?

Probably a big component of that slide is testing older OS versions of older hardware with very conservative estimates, but another big part is that a lot of these are system tests that place a lot of emphasis on RAM speed, storage speed, and so on. And that nVidia probably used a tweaked up reference tablet (or board) with fast DDR3 vs some phone that may not have even been allowed to run at full clocks.

I for one feel like starting the graphs away from 0 doesn't even begin to approach the marketing lies nVidia puts into their presentations. One is massaging how you present the data in an easy to recognize fashion to give an illusion of a bigger impact on otherwise truthfully reported material. Compare that with presenting data that is either taken under completely unfair testing conditions or just utterly made up.. it's no contest.

 

[ams]

And "only" 26% is actually kind of a lot.

26% sounds like a lot, but the T4 GPU performance in GLBenchmark 2.5 Egypt HD Offscreen (1080p) is about 36% higher in comparison, so the GPU perf/MHz is actually higher with T4.

When I look at Tegra 4's GPU I see some very strange balancing choices. Compared to Tegra 3 they've increased both fragment and vertex processing capability by 8x, yet have only increased throughput of the triangle engine (setup, clipping, etc) by 1.5x. I can't begin to explain the rationale behind this one.

You would have to ask NVIDIA about that, but most likely they had to make some hard choices in order to increase performance while keeping die size under control. The GLB 2.5 Egypt HD tests do not appear to be particularly sensitive to increases in triangle throughput.

And while nVidia might have much more efficient Cortex-A15s on TSMC 28nm than Samsung's 32nm ones I kind of doubt it.

Is there not a significant difference in transistor density and power consumption between 32nm @ Samsung and 28nm @ TSMC? Anyway, the fact remains that we don't have much in the way of application specific power consumption measurements with Cortex A15 on TSMC's 28nm fabrication process.

Now that slide, that's just grade A bullshit. Just compare the number nVidia gives for Sunspider on Tegra 4, 499ms, with what a 1.9GHz S600 gets: http://www.gsmarena.com/samsung_galaxy_s4_vs_htc_one-review-913p6.php

Difference is only 1.62x, so where does 2.5x come from, aside from nVidia's imagination?

Clearly Samsung has done a lot of tuning with Sunspider, because even the old SGS3 is as fast as the new HTC One! Anyway, compared to the S600 in the HTC One, T4 is more than 2.2x faster. From Anandtech's measurements, you can see that the S600 in the HTC One is consistently only ~ 30% faster than the S4 Pro in Sunspider, Google Octane, Kraken, Vellamo HTML5, and Vellamo Metal, and 0% faster in BrowserMark 2.0:

http://images.anandtech.com/graphs/graph6747/53584.png
http://images.anandtech.com/graphs/graph6747/53586.png
http://images.anandtech.com/graphs/graph6747/53590.png
http://images.anandtech.com/graphs/graph6747/53600.png
http://images.anandtech.com/graphs/graph6747/53601.png
http://images.anandtech.com/graphs/graph6747/53650.png

And we haven't even begun to talk about Web Page Load, SPECInt2000, WebGL Aquarium, etc.

To deny that quad-core Cortex A15 will be way faster in most of these tests compared to S600 in insane. Unless you can prove that quad-core Cortex A15 at 28nm consistently has more than 1.5-3x higher power consumption when completing these benchmarks compared to quad-core S600, then you really have no way of knowing whether or not there is a big difference in perf/w.

 

[Exophase]

Sunspider is an awful benchmark yes, but that doesn't make it okay to make up numbers for it.

Samsung has done a lot of tuning with it? You have real evidence for that? What do you even have in mind? Samsung made huge improvements to V8 in their phones that they're hording? And why does Optimus G Pro, also S600, score similarly to Galaxy S4?

As for your claim that HTC One is consistently only 30% faster vs S4.. I'm going to go ahead and give the best like vs like I can and compare it with the S4 HTC One X. Afterall, how can you claim that a platform is unfairly boosting S600's scores but then use the best S4 numbers? Why would platform-specific boosting not apply there? This is what you get:

Sunspider: 45%
Octane: 59%
Kraken: 55%
Vellamo HTML5: 62%
Vellamo Metal: 43%

Definitely not consistently 30% faster.

And yes, I'm leaving out benchmarks even partially influenced by GPU performance since that has no place in a CPU performance comparison.

When scores vary a lot using the same SoC in different phones what does that tell you? You can't just use whichever one best illustrates your case.

 

[ams]

Samsung has done a lot of tuning with it? You have real evidence for that?

Look at the data. Do you think it makes sense that the GS3 (with quad-core Cortex A9 @ 1.4GHz) has the same Sunspider score as the HTC One (with quad-core Krait 300 @ 1.7GHz) if there was no tuning? Do you think it makes sense that the GS4 (with quad-core Krait 300 @ 1.9GHz) has a 38% higher Sunspider score than the HTC One (with quad-core Krait 300 @ 1.7GHz), even though the difference in operating frequency is only ~ 12%? No other benchmark shows anywhere near that level of difference when comparing GS4 to HTC One (and ironically enough, the HTC One actually beats the GS4 in the Vellamo HTML5 test).

As for your claim that HTC One is consistently only 30% faster vs S4.

Read what I wrote more carefully. I said that the HTC One is consistently only ~ 30% faster than S4 Pro. You can clearly see ~ 30% performance delta between S4 Pro and S600 if you compare the HTC One to the Nexus 4 or Droid DNA in Anandtech's benchmarks linked above. The GS4 has a slight performance advantage on average compared to the HTC One primarily due to the difference in operating frequencies.

 

[Exophase]

Right, so if Samsung has magic for their Sunspider scores why does LG get the same thing with S600? And more importantly, why do you think this advantage wouldn't apply to nVidia's scores? And how are you certain that an advantage in one Javascript benchmark doesn't translate to an advantage in a bunch of other Javascript benchmarks, which is what the bulk of these scores are for?

No, I'm not denying that there's a software issue here. I'm saying maybe it's more of HTC having a problem than everyone else having an advantage. Kind of puts a different dimension on these cross-vendor comparisons, don't you think?

 

[ams]

The LG Optimus G Pro (with S600) in Chrome has virtually the same Sunspider score as the HTC One in Chrome. So clearly LG and Samsung have done some Sunspider/JS optimization in their stock browser.

 

[Turbotab]

I personally think Sunspider is a terrible benchmark, just to show how variable it is, a Snapdragon 600 (1.7 GHz) equipped Asus Padfone Infinity scores ~ 875ms

http://www.gsmarena.com/asus_mwc_2013-review-899p2.php

Also, what is the point of comparing benchmarks from Nvidia's CES T4 reference tablet, that was obviously enjoying a far more generous power envelope than likely in shipping devices?

At CES T4 scored 36,000 in AnTuTu, whereas leaked scores for a Toshiba AT10LE-A, which is meant to use T4, only scores 28,000. We all know that in unconstrained power environments that A15 will outscore Krait 300, but constrained to the limits of shipping devices that lead is not so dominate.

http://gsminsider.com/2013/04/15/at10le-a-antutu-benchmark-result/

 

[Exophase]

So LG and Samsung (and ASUS, thanks TurboTab) either both independently optimized the stock browser or are colluding to perform improvements together. Or maybe they're just both using a newer version of what's publicly available from Google and HTC isn't. I don't really know exactly what's going on but the basic point remains - I have no reason to believe nVidia wouldn't use the best JS engine code they can for their tests while not affording the same benefit to the competition. And I have no reason to believe they would be dramatically wrong on their Sunspider comparison but dead on on all the other Javascript benches which is a big chunk of the slide. Yes, Sunspider is awful but so are most benchmarks used on phones, even if they're not all quite as awful. Some that seem like they have a decent set of tests, like Geekbench, end up being awful due to being incredibly erratic for no known reason.

And for things they had to compile themselves like SPECInt2k we have no assurance that they're using the same binaries nor do we have any idea how to reproduce the scores (although other sources have shown them to be on the level). The remaining native-binary scores show some of the smaller gains. And are probably estimated using a 1.7GHz S600, of course without specifying that. And are probably using a platform with a bunch of advantages that have nothing to do with the SoC.

I get it, shitty marketing is everywhere, but that doesn't make it fair game for adding to a technical argument. The idea that there's this prominent 2-3x IPC advantage in Tegra 4's Cortex-A15 vs S600's Krait doesn't match up with a whole load of real world information we have on those chips.

 

[ams]

At the end of the day, application performance matters, and the quad-core Cortex A15 in T4 appears to be way ahead of the quad-core Krait 300 in S600 with a wide variety of applications including but not limited to: SPECInt2000, Sunspider, Web Page Load, WebGL Aquarium, Google Octane, Geekbench, Kraken, Vellamo HTML5, Vellamo Metal, etc. Whether or not the Sunspider difference is 1.7-2.2x in actuality vs. 2.3x in the slide doesn't really matter. The performance difference is almost staggering, and is anywhere between 1.5-3x higher for the A15's virtually across the board with most of these applications. So again, unless the A15's on 28nm actually consume more than 1.5-3x more power than Krait 300 to run these apps, the perf/w will be competitive and the overall average power consumption will be competitive too, at the expense of higher peak power.

 

[Ailuros]

26% sounds like a lot, but the T4 GPU performance in GLBenchmark 2.5 Egypt HD Offscreen (1080p) is about 36% higher in comparison, so the GPU perf/MHz is actually higher with T4.

While comparing a tablet to a smartphone SoC, I'm sure there's a fair metric for it to talk about perf/MHz which must be a recent invention also. If T4 makes it into a smartphone it'll then be a good comparison to the S4 under its final real frequencies and not what is supposed to end up in a 5W TDP tablet.

You would have to ask NVIDIA about that, but most likely they had to make some hard choices in order to increase performance while keeping die size under control. The GLB 2.5 Egypt HD tests do not appear to be particularly sensitive to increases in triangle throughput.

And how much would it had cost them in hw to increase the trisetup unit by say just 50% in throughput? Why do you even bring up synthetic benchmark results when someone makes a legitimate point? The ALU lane increase from T3 to T4 is in terms of unit count at 6x times. If you leave the trisetup part of a GPU untouched and it'll only increase as much as core frequency it's inevitable that you end up with a question mark what the rather big increase in VS ALUs is actually good for.

-----------------------------------------------------------------------------------------------------------------------

Small flashback to die area claims:

http://www.anandtech.com/show/5072/nvidias-tegra-3-launched-architecture-revealed

Tegra2

Tegra3

We have a total die area for Tegra2 at 49mm2@40nm TSMC and Tegra3 at 80mm2@40nm TSMC. It shouldn't be too hard to estimate from those how big each ULP GF block is on those. As a reminder:

T2 ULP GF = 1 Vec4 PS, 1 Vec4 VS, 2 TMUs, peak 333MHz
T3 ULP GF = 2 Vec4 PS, 1 Vec4 VS, 2 TMUs, peak 520MHz
T4 ULP GF = 12 Vec4 PS, 6 Vec4 VS, 4 TMUs, peak 672MHz

Now just from 40 to 28nm at TSMC with a crapload of more units for the T4 GPU and another increase in frequency and the T4 ULP GF block is still at just 10.5mm2? What exactly am I missing?

 

[Exophase]

Whether or not the Sunspider difference is 1.7-2.2x in actuality vs. 2.3x in the slide doesn't really matter.

Whether it's 1.7x or 2.3x does matter. You can't say things like this then hammer on again all the figures in the slide. You shouldn't keep including a bunch of GPU benchmarks in a discussion of CPU perf/W (and the data for making any kind of GPU perf/W comparison is very limited). Saying that Tegra 4 will tend to have a 1.5-3x IPC advantage vs S600 is a slap in the face to anyone who cares about talking about CPU design implications.

It seems like you're willing to take any opportunity to make nVidia's products look as good as possible vs everything else but I'm interested in real technical discussions that are about something more than which brand wins. And that means spending some time reasoning about expected outcomes based on everything we've seen thus far (Cortex-A9 vs Krait, Cortex-A9 vs Cortex-A15, S4/Pro's Krait vs S600's, etc) and known technical details of the uarchs and SoCs implementing them.

No we don't have power measurements for Tegra 4 but we have numbers nVidia claim. Let me see if I'm getting this right - about 670mW @ 825MHz (according to you NOT for the battery saver core, according to Arun there's a good chance it'll use more power at that clock speed anyway). This is using nVidia's numbers same as the performance claims, but at least they're absolute and less subject to fudging.

We can compare to a 1.5GHz Krait core in the first S4s on a very early TSMC 28nm release (without HKMG!) using 750mW, as measured by Anandtech. Krait 400 processors are going to be on a much more mature process and using HKMG, which will definitely give them some perf/W advantage and put them on equal footing with nVidia.

You probably think that at 845MHz Tegra 4's CPU is a slam dunk vs Krait 200 @ 1.5GHz. I doubt this. Because in my experience you rarely see IPC improvements anywhere close to this going from Samsung's A9s to A15. You can see comparisons on Phoronix, you can also see a bunch here http://marcin.juszkiewicz.com.pl/2012/11/28/lets-compare-some-cpu/ and there's the usual http://www.7-cpu.com/ Clock scaling won't be perfect (if you assume it the numbers look pretty bad) but it's easy to see there is often not a huge IPC boost for these tests. This is all pretty computationally biased and I'm not saying it's hugely representative but it gives you a starting point against claiming that there's ALWAYS going to be this big (at least 1.5x!) IPC win.

I can also say that in my own tests for emulator performance I've seen IPC delta of around 1.4x to 2.1x going from a Cortex-A8 SoC @ 600MHz (OMAP3530) to Exynos 5250 @ 1.7GHz, w/perfect scaling assumed. Consider that A9 tends to give around 20-30% IPC improvement vs Cortex-A8 and you get something like 1.08x to 1.75x IPC boost. For reference, ARM themselves claimed 1.4x for Dhrystone, 1.5x for "integer" (one would figure SPECInt2k), 2x for "floating point" (SPECFp2k?) and 2x for "memory." They also said 1.5x was their general IPC goal. I have yet to see an application that offers anywhere close to 3x an IPC boost from Cortex-A9.

You usually saw the first Kraits outperforming Cortex-A9s at the same clock when running the same software. Sometimes only very modestly, but this is what reviews usually showed. We're now looking at Kraits that have a fair bit more IPC than the first ones. So if saying 1.5x to 3x IPC improvement from Cortex-A9 to Cortex-A15 sounds silly then saying it for Krait 400 sounds really silly. And I refuse to base an entire technical argument around this premise because some nVidia slide says so.

I eagerly wait real information and testing, both power and performance, and I hope we get some Qualcomm boards that can run Linux so we can run real benchmarks using the same binaries under controlled conditions. But given the information we have thus far I'm not holding my breath that nVidia will offer similar or better perf/W on the CPU vs Krait. Rough impression based on all the numbers I've presented here is an nVidia Tegra 4 at 800-900MHz will use about the same power as a Krait core at let's say 1.4-1.5GHz, while having average performance under same Krait at 1.2GHz (max 50% better IPC boost average over many applications under fair testing conditions meaning same OS, same settings, same support hardware like memory and storage). I also expect 800-900MHz to be around the sweet spot for nVidia; it'll possibly continue to scale more favorably at lower MHz but almost certainly less favorably at higher.

 

[ams]

Whether it's 1.7x or 2.3x does matter.

Look, this isn't rocket science. The slide clearly estimated S600 performance by adding 30% to the S4 Pro measured performance. So the data in the slide appears to be reasonably in line with the S600 performance of the HTC One. Yes, the Samsung S600 variant is slightly faster, but it is only slightly ahead of the HTC variant across most applications and still way behind T4 across most applications.

Saying that Tegra 4 will tend to have a 1.5-3x IPC advantage vs S600 is a slap in the face to anyone who cares about talking about CPU design implications.

I never said "IPC" advantage, I said "application performance" advantage, get it? Application performance matters, period. You claim that you want to have a technical discussion, but then you ignore one application data point after another after another. Is SPECInt2000, Sunspider, Web Page Load, WebGL Aquarium, Google Octane, Geekbench, Kraken, Vellamo HTML5, Vellamo Metal, not good enough for you? If so, then please suggest something that is, and then hope and pray that someone will measure it for you.

No we don't have power measurements for Tegra 4 but we have numbers nVidia claim. Let me see if I'm getting this right - about 670mW @ 825MHz (according to you NOT for the battery saver core, according to Arun there's a good chance it'll use more power at that clock speed anyway). This is using nVidia's numbers same as the performance claims, but at least they're absolute and less subject to fudging.

We can compare to a 1.5GHz Krait core in the first S4s on a very early TSMC 28nm release (without HKMG!) using 750mW, as measured by Anandtech. Krait 400 processors are going to be on a much more mature process and using HKMG, which will definitely give them some perf/W advantage and put them on equal footing with nVidia.

I'm not sure what you are talking about. The NVIDIA power consumption data was at one specific relatively low frequency using one main CPU core with SPECInt. Anand never measured anything of the sort on Krait.

The mathematics here are very simple. If the quad-core Cortex A15 in T4 has 1.5-3x performance advantage across a variety of applications, then it will need to have even more than 1.5-3x power consumed in order to have worse perf/w on these same applications compared to Krait 300 in S600. Hopefully Anand will be able to accurately measure app-specific power consumption on these SoC's to put this issue to rest once and for all.

 

[Ailuros]

Nah none of us would know what we're talking about nor can any of us detect bullshit when presented in a marketing slide from any IHV. I don't expect to see reasonable answers or disputes to any points that are going to appear, just the usual baseless parotting of marketing hyperboles.

 

[Exophase]

No, you didn't say IPC advantage.. it just so happens S600 is out in the wild running at 1.9GHz and Tegra 4 is running at 1.9GHz so peak performance advantage = IPC advantage. I thought that would be clear enough without me having to spell it out. Or is it somehow okay for us to perform this performance comparison using an S600 clocked at less than a Tegra 4 will hopefully clock at at launch (yes, there's a history of SoCs clocking lower than the manufacturer said they would)?

The 670mW number comes from Linley's report on Tegra 4 which I presume you already read but here it is if you haven't: http://www.slideshare.net/fullscreen/caulfield2600/tegra-4-outperforms-snapdragon-17176163/1. The 750mW number came from very early reviews Anandtech did for S4: http://www.anandtech.com/show/5559/...mance-preview-msm8960-adreno-225-benchmarks/4

There's more data here: http://www.anandtech.com/show/6536/arm-vs-x86-the-real-showdown/4 This CPU should be capable of clocking up to 1.7GHz and the power looks to be peaking around 600mW average for loaded single threaded tests, with some small spikes up to about 800mW. That would seem to be in line with my expectations that newer revisions of the same SoC (hence, higher peak clock allowed) benefit from process maturation. And figures because Qualcomm themselves estimated much higher initial clocks than 1.5GHz. But I'm pretty sure this is still lacking HKMG which S600 and S800 should not be, nor should Tegra 4.

I don't care how legitimate nVidia's methodology in estimating scores was or wasn't. I care about the relevance this slide has on this discussion. Not really sure where you get off saying you're talking application performance that I'm ignoring data point after data point. Reality is a) nVidia is using benchmarks just as much as I am, many synthetic ones, very little "application performance" b) testing details are vague to non-existent c) nVidia is a highly partial party with a history of presenting crooked benchmark data d) the data is known to be at least on some points flat out inconsistent with what third parties have reported. And you have to ask why this is not good enough for me? You object when I prefer tests on controlled platforms by impartial third parties with typically much more specific environmental information than some marketing slide? Blowing these concerns off with things like "who cares if it's really 1.7x or 2.3x?" makes me wonder if you're really taking this seriously.

I don't have to hope and pray that anyone measures something for me because lots of people already have, I've presented plenty comparing Cortex-A9 and Cortex-A15. If you don't think that that envelopes Krait 400 vs Cortex-A15 then you are disconnected from reality but if I really have to make a case that it'll have an average IPC advantage vs (non-R4, mind you!) Cortex-A9s then I'm sure I can do that. I just don't want to go through the trouble of spelling out something that's common knowledge.

If nVidia had some big software advantage like Intel has enjoyed I would concede that this has a bearing on a real world perf/W discussion. But they don't, they're using the same compilers and same operating systems and system software.

And in all this I guess you must also think Samsung has no problem launching two variants of a product with the same name but some 1.26x to 2.5x difference in CPU performance (adjusted from 1.9GHz for Tegra 4 to 1.6GHz for Exynos 5410, as usual more optimistic scaling than realistic) OR maybe you think nVidia has a big IPC advantage in their Cortex-A15 vs Samsung's.

 

[Ailuros]

And in all this I guess you must also think Samsung has no problem launching two variants of a product with the same name but some 1.26x to 2.5x difference in CPU performance (adjusted from 1.9GHz for Tegra 4 to 1.6GHz for Exynos 5410, as usual more optimistic scaling than realistic) OR maybe you think nVidia has a big IPC advantage in their Cortex-A15 vs Samsung's.

When I said elsewhere that I have the gut feeling that Samsung adjusted frequencies for the S600 and 5410 in order for the performance ballpark not to have a noticably large gap you found the notion strange for Samsung. From what I've been able to gather so far the gap between the two (for 1890/480MHz vs. 1600/1200/480MHz) is usually within the 10% margin with the 5410 though probably having the tendency of slightly better battery life. IMHO the GPU in the S600/S4 is definitely clocked higher than in the Nexus4 and the 544MP3 in the 5410 didn't peak at the probably originally projected 533MHz frequency.

 

[Exophase]

I didn't find the notion of Samsung adjusting frequencies to match performance particularly strange, but I didn't think that such a thing would really be warranted. What I found strange was the notion of Samsung needing to set Cortex-A15 at 1.2GHz to match Krait at 1.9GHz exaggerated (and of course it has been demonstrated that they didn't do that). And still find that notion strange. But that ratio seems downright modest compared to what nVidia is saying and ams is supporting...

Now, does that mean I think that the Exynos S4s are held at 1.6GHz instead of 1.8GHz for this reason? I doubt it, simply because I think 4x1.8GHz pushes the allowable power consumption too high for a phone - way past killing your battery and into thermal limits. Samsung could have allowed higher clocks with only one or two cores active but they never did before so I don't think they'd start now unless the limitations were dire (like the original 1.2GHz claim)

Do you have any links to the performance data you're looking at? 10% would mean 30% better IPC (with perfect scaling, blah blah blah usual disclaimer), which sounds about right to me. First I heard about 1890MHz instead of 1900, too small to make a difference on the discussion but still kind of weird!

 

[Turbotab]

I didn't find the notion of Samsung adjusting frequencies to match performance particularly strange, but I didn't think that such a thing would really be warranted. What I found strange was the notion of Samsung needing to set Cortex-A15 at 1.2GHz to match Krait at 1.9GHz exaggerated (and of course it has been demonstrated that they didn't do that). And still find that notion strange. But that ratio seems downright modest compared to what nVidia is saying and ams is supporting...

Now, does that mean I think that the Exynos S4s are held at 1.6GHz instead of 1.8GHz for this reason? I doubt it, simply because I think 4x1.8GHz pushes the allowable power consumption too high for a phone - way past killing your battery and into thermal limits. Samsung could have allowed higher clocks with only one or two cores active but they never did before so I don't think they'd start now unless the limitations were dire (like the original 1.2GHz claim)

Do you have any links to the performance data you're looking at? 10% would mean 30% better IPC (with perfect scaling, blah blah blah usual disclaimer), which sounds about right to me. First I heard about 1890MHz instead of 1900, too small to make a difference on the discussion but still kind of weird!

If you're interested, this the "2000 MHz" variant of the frequency / voltage table found in the HTC One's kernel source release.

static struct acpu_level tbl_PVS5_2000MHz[] __initdata = {
{ 1, { 384000, PLL_8, 0, 0x00 }, L2(0), 875000 },
{ 1, { 486000, HFPLL, 2, 0x24 }, L2(5), 875000 },
{ 1, { 594000, HFPLL, 1, 0x16 }, L2(5), 875000 },
{ 1, { 702000, HFPLL, 1, 0x1A }, L2(5), 875000 },
{ 1, { 810000, HFPLL, 1, 0x1E }, L2(5), 887500 },
{ 1, { 918000, HFPLL, 1, 0x22 }, L2(5), 900000 },
{ 1, { 1026000, HFPLL, 1, 0x26 }, L2(5), 925000 },
{ 1, { 1134000, HFPLL, 1, 0x2A }, L2(15), 937500 },
{ 1, { 1242000, HFPLL, 1, 0x2E }, L2(15), 950000 },
{ 1, { 1350000, HFPLL, 1, 0x32 }, L2(15), 962500 },
{ 1, { 1458000, HFPLL, 1, 0x36 }, L2(15), 987500 },
{ 1, { 1566000, HFPLL, 1, 0x3A }, L2(15), 1012500 },
{ 1, { 1674000, HFPLL, 1, 0x3E }, L2(15), 1050000 },
{ 1, { 1782000, HFPLL, 1, 0x42 }, L2(15), 1087500 },
{ 1, { 1890000, HFPLL, 1, 0x46 }, L2(15), 1125000 },
{ 0, { 0 } }

 

[french toast]

I personally think Sunspider is a terrible benchmark, just to show how variable it is, a Snapdragon 600 (1.7 GHz) equipped Asus Padfone Infinity scores ~ 875ms

http://www.gsmarena.com/asus_mwc_2013-review-899p2.php

Also, what is the point of comparing benchmarks from Nvidia's CES T4 reference tablet, that was obviously enjoying a far more generous power envelope than likely in shipping devices?

At CES T4 scored 36,000 in AnTuTu, whereas leaked scores for a Toshiba AT10LE-A, which is meant to use T4, only scores 28,000. We all know that in unconstrained power environments that A15 will outscore Krait 300, but constrained to the limits of shipping devices that lead is not so dominate.

http://gsminsider.com/2013/04/15/at10le-a-antutu-benchmark-result/

875ms for KRAIT..wow.

As you are pointing out and many here seem to be missing..tegra 4 reference tablet has a very high tdp headroom, ddr3 1866? And likely other enhancements I cant remember (I mentioned these a few pages back but cant remember the exact details)

The dodgy nvidia marketing graphs compare that platform against qualcomms rough edtimates for its smartphone platform...we know qualcomm is ultra realistic with its performance projections(maybe understates). .and nvidia drinks the kool aid

In a smartphone then...you can completely discount any of that unrealistic drivel fron nvidia ever coming close to being accurate.

Exynos 5410 according to leaks...gets slightly better benchmarks than s600 gs4 running the same software. .and that carries cortex A7s to balance power consumption. ..tegra 4 has no such benefit.

Tegra 4 is a tablet oriented SOC...snapdragon more a smartphone SOC.

I think the fact tegra 4i is set to launch soon after maybe points to this.

 

[Ailuros]

I didn't find the notion of Samsung adjusting frequencies to match performance particularly strange, but I didn't think that such a thing would really be warranted. What I found strange was the notion of Samsung needing to set Cortex-A15 at 1.2GHz to match Krait at 1.9GHz exaggerated (and of course it has been demonstrated that they didn't do that). And still find that notion strange. But that ratio seems downright modest compared to what nVidia is saying and ams is supporting...

Now, does that mean I think that the Exynos S4s are held at 1.6GHz instead of 1.8GHz for this reason? I doubt it, simply because I think 4x1.8GHz pushes the allowable power consumption too high for a phone - way past killing your battery and into thermal limits. Samsung could have allowed higher clocks with only one or two cores active but they never did before so I don't think they'd start now unless the limitations were dire (like the original 1.2GHz claim)

Do you have any links to the performance data you're looking at? 10% would mean 30% better IPC (with perfect scaling, blah blah blah usual disclaimer), which sounds about right to me. First I heard about 1890MHz instead of 1900, too small to make a difference on the discussion but still kind of weird!

Kishonti is listing 1890MHz which I had mentioned somewhere here and you might recall it. The 10% is so far collective hearsay from various sides and it would be best if we'd treat it as such until reliable 3rd party benchmarks/analysis arrives and we can read into it. I actually wanted to fetch a 5410/S4 since a coworker has already a S600/S4 to run them side by side, but in my neck of woods only the latter variant is available for the time being.