NVIDIA Tegra Architecture
- Thread starter french toast
- Start date
-
- Tags
- nvidia
french toast
Veteran
Simple...it strikes a perfect balance of performance/mm2..with likely very competitive power consumption.
I think it should be somewhere in the krait 100 ballpark on integer clock for clock..just a guesstimate on my part.
This would make a perfect midrange smartphone platform. ...
Last edited by a moderator:
french toast
Veteran
Yea precisely.
What was I going on about with nvidias disingenuous marketing before?...
#shambles.
I'm breaking my head over that slide, how the hell they convinced themselves to integrate A15s in Tegra4 SoCs and not higher clocked A9 cores instead
This slide from NVIDIA is looking specifically at SPECInt IPC. The Tegra 4/4i data is from NVIDIA R&D, but I'm not sure where the Snapdragon 800 data is from (apparently some publically available source of data according to the slide footnotes). According to Anandtech, "NVIDIA has noted that it using R4 of ARM’s Cortex A9, which includes higher IPC thanks to the addition of a better data prefetching engine, dedicated hardware for cache preload instructions and some larger buffers". That said, the SPECInt IPC for Cortex A15 is still about 50% higher than R4 Cortex A9.
That's great, but having to wait all the way until the end of 2013 (or early 2014) is not so great
D
Deleted member 13524
Guest
With those release schedules, nVidia just put themselves out of the game for 2013 and 2014.
Tegra 3 was available in Q4 2011, and it will take them over two years to release a successor for smartphones. That's not good at all, IMO.
Tegra 3 was available in Q4 2011, and it will take them over two years to release a successor for smartphones. That's not good at all, IMO.
Tegra3 as all Tegras where only ever just one SoC serving both tablets and what NV calls superphones or let me translate the latter into high end smartphones. T4 is aiming to serve those exact two markets as all its predecessors and Grey/T4i is aimed to serve the mainstream smartphone market which is a novelty so far for NVIDIA. So far they couldn't afford to work on two different SoCs. Grey must have been conceived later as an idea than T4 IMHO and from what I recall hearing in the past Grey was developed by their team in India.
I'm not arguing about the late/early part of the story, it's just that Grey is a novel SoC animal for NV's roadmap. If you'd want to criticise NVIDIA one good starting point would be why the heck they don't invest more resources in an as cut throat market as the smartphone/tablet market turned out to be.
....or maybe revised schedule: Grey now and T4 at the end of 2013 would have been better imho.
I was so annoyed at this Nvidia PR slide, stating that Snapdragon S800 will be creamed by both T4 and the newly announced A9 based T4i, that I went from lurking to being a member!
http://androidandme.wpengine.netdna-cdn.com/wp-content/uploads/2013/02/tegra-4i-vs-krait-s800.jpg
Ignoring the fact that the S800 figures are their estimates, something they only barely allude to, as they clearly won't have hard data, unless they are guilty of spying What amazes me is that they seem to quote figures for a S4 @ 1.5 GHz. In the benchmarks below, they seem to multiple the single threaded SpecINT score by the numbers of cores in the respective SoC.
"Solution A" is a dual-core Krait @ 1.5 GHz and scores 1517 /2 = 758.5 per core.This figure is earily close to one quoted by Nvidia under Snapdragon 800.
758.5 / 1500 = 0.506 per MHz
Even ignoring any architectural improvements that Qualcomm have stated for their new Krait 400 cores, a 2.3 GHz S800 should be scoring at least (0.506 x 2300) 1163.6 in SpecINT, virtually the same as the A15 Tegra 4 @ 1.9 GHz. If Qualcomm are true to their word, we could even see figures 15% ~ higher than my first prediction, which well beyond T4.
I have spent plenty of money at the altar of Nvidia, I don't mind fighting talk, but making up such spurious rubbish insults us all.
www.inpai.com.cn/doc/phone/187065.htm
http://androidandme.wpengine.netdna-cdn.com/wp-content/uploads/2013/02/tegra-4i-vs-krait-s800.jpg
Ignoring the fact that the S800 figures are their estimates, something they only barely allude to, as they clearly won't have hard data, unless they are guilty of spying
What amazes me is that they seem to quote figures for a S4 @ 1.5 GHz. In the benchmarks below, they seem to multiple the single threaded SpecINT score by the numbers of cores in the respective SoC."Solution A" is a dual-core Krait @ 1.5 GHz and scores 1517 /2 = 758.5 per core.This figure is earily close to one quoted by Nvidia under Snapdragon 800.
758.5 / 1500 = 0.506 per MHz
Even ignoring any architectural improvements that Qualcomm have stated for their new Krait 400 cores, a 2.3 GHz S800 should be scoring at least (0.506 x 2300) 1163.6 in SpecINT, virtually the same as the A15 Tegra 4 @ 1.9 GHz. If Qualcomm are true to their word, we could even see figures 15% ~ higher than my first prediction, which well beyond T4.
I have spent plenty of money at the altar of Nvidia, I don't mind fighting talk, but making up such spurious rubbish insults us all.
www.inpai.com.cn/doc/phone/187065.htm
What Anandtech said is that Tegra 4i (with integrated Icera i500 baseband processor, R4 Cortex A9 CPU, and 60 core ULP Geforce GPU) will not show up in commercial smartphones until end of 2013 (or early 2014). The expectation is that Tegra 4 (with separate Icera i500 baseband processor, Cortex A15 CPU, and 72 core ULP Geforce GPU) will show up much earlier than that, considering that Tegra 4 and i500 have been sampling for some time now.
Yes, the R4 Cortex A9 appears to give better performance per mm^2 than Cortex A15 based on SPECInt (although performance per watt differences are unknown), but the absolute performance of quad Cortex A15 will still be better than quad R4 Cortex A9, and it wouldn't be sensible to use eight R4 Cortex A9 cores as an alternative to the quad Cortex A15.
"Next gen" is just a word, it does not mean anything
. Actually these results are quite good for Krait, because single thread apq8064 score is 430-450(I've measured it with locked clocks), in perfect case 2.3 Ghz apq8064 would have 450*1.53 = 690 and if assume a 10% ipc grow for second generation Krait you will get - 690*1.1 = 759 score
"Next gen" is just a word, it does not mean anything
What's your point exactly? I'm referring to the revision with the ~10-15% IPC bump just like you are, what was I supposed to call it instead?
But a SPECInt score is not a measurement solely of hardware since it depends heavily on the compiler and libraries. Who knows what compiler differences are present in nVidia's test, but they could have actually used the same binary on both (and adjusted the Krait score like you did) but probably didn't. I don't believe for a minute that even this newer revision A9 has 20% higher perf/MHz than the revised Krait, specifically in SPECInt (2K?), under a fair comparison.
We all know nVidia isn't above using unfair comparisons with very different compilers, this time I guess they just aren't reporting it..
I am quite sure they are using the same binaries for all arm v7 devices, it's a common practice
Do you have a Tegra 3 or Exynos 5 device you can run it on to see if the scores are anywhere close?
Let's see how the Qualcomm will react, they will reply fast if nvidia scores are bullshit, otherwise they will be silent about SPEC
Similar threads
- Locked
