NVIDIA's Tegra 3 Launched: Architecture Revealed
by Anand Lal Shimpi on November 9, 2011 12:34 AM ESTASUS' Transformer Prime: The First Tegra 3 Tablet
With Tegra 2, Motorola was the primary launch partner both for smartphones and tablets. Since then, ASUS has risen in the ranks and is now a serious competitor in the Android tablet space. It's no surprise that the first Tegra 3 tablet out of the gate is ASUS' Transformer Prime.
ASUS will launch the Transformer Prime in the US before the end of the year. The tablet's specs are below:
| Tablet Specification Comparison | ||||||
| ASUS Eee Pad Transformer | ASUS Eee Pad Transformer Prime | Apple iPad 2 | Samsung Galaxy Tab 10.1 | |||
| Dimensions | 271mm x 175mm x 12.95mm | 263 x 180.8 x 8.3mm | 241.2 x 185.7 x 8.8mm | 256.6 x 172.9 x 8.6mm | ||
| Display | 10.1-inch 1280 x 800 | 10.1-inch 1280 x 800 Super IPS+ | 9.7-inch 1024 x 768 IPS | 10.1-inch 1280 x 800 PLS | ||
| Weight | 675g | 586g | 601g | 565g | ||
| Processor | 1GHz NVIDIA Tegra 2 (2 x Cortex A9) | 1.3GHz NVIDIA Tegra 3 (4 x Cortex A9) | 1GHz Apple A5 (2 x Cortex A9) | 1GHz NVIDIA Tegra 2 (2 x Cortex A9) | ||
| Memory | 1GB | 1GB | 512MB | 1GB | ||
| Storage | 16GB + microSD card | 32GB/64GB + microSD slot | 16GB | 16GB | ||
| Pricing | $399 | $499/$599 | $499 | $499 | ||
Final Words
At a high level Tegra 3 doesn't surprise us much. The improved GeForce GPU should deliver tangible performance gains both through increased operating frequency and more pixel shader hardware. CPU performance should also be better than Tegra 2 based designs thanks to an increase in clock speed, the inclusion of MPE and the availability of more cores for threaded applications. In the move from one to two cores we saw significant performance increases across the board in Android. I don't expect that we'll see gains of a similar magnitude in moving from two to four cores, but there will be some benefit.
For the majority of use cases I believe NVIDIA has done the hardware homework necessary to extend battery life. Individual cores can now be power gated and the companion core should do most of the lifting while your device is locked or mostly idle, processing background tasks.
How much of an impact we'll actually see from all of this remains to be seen. We hope to have our hands on the first Tegra 3 hardware in the coming weeks, so before the year is up we'll hopefully have some answers.
Facebook
Twitter
RSS
94 Comments
View All Comments
psychobriggsy - Friday, November 11, 2011 - link
In the Android market, it really doesn't matter what features Apple includes in their in-house SoC for their iOS devices.Considering that manufacturers are having problems fabbing larger high-DPI displays, I also wouldn't be betting on the iPad 3 having a higher resolution display. And Apple would go for 2048x1536 for simplicities sake.
Five months is also a long time in the ARM SoC market, one that NVIDIA will try to make use of. Let's just hope the product meets the hype when reviews roll in.
name99 - Wednesday, November 9, 2011 - link
I've asked this before, and I will ask it again:What software on Android, shipping TODAY, is capable of using 4 cores usefully?
The browser? The PDF viewer? Google Earth? If so, they're all ahead of their desktop cousins.
Yes, yes, people are buying the future. And, sure, one day, software will be revved to use 4 cores. (But, this being Android, chances are, the particular device you buy this year using Tegra3 will NOT be revved.)
I'm not trying to be snarky here, just realistic. It seems to me the competitors ARM manufacturers are targeting the real world, where dual cores can (to some extent) usefully be used. But nVidia is requiring people who adopt this chip to pay for power that, realistically, they're not going to use. This seems a foolish design choice. It seems to me far more sensible for mobile to basically track (lagging by about a year) desktop. Desktop is seeing quad-core adoption in a few places, but it's hardly mainstream --- and I'd say that until, let's say, the low-end MacBook Air is using quad core, that's an indication that "software" (as a general class) probably hasn't been threaded enough to make quad-core worth-while in mobile.
Yes, it's harder, but until then, I'd say far more useful to look at what's ACTUALLY causing people slowness and hassle on phones and tablets, and add THAT to your chips. So, faster single-threaded core --- great. But think more generally.
Flash on these devices is still slow. Could you speed it up somehow --- maybe a compression engine to transparently compress data sent to/from flash? Likewise app launch is slow. Are there instructions that could be added to speed up dynamic linking? Memory is a problem, and again transparent compression might be helpful there.
Basically --- solve the problems people actually have, even if they are hard, NOT the problems you wish people had because you know how to solve those.
psychobriggsy - Friday, November 11, 2011 - link
The video shows web browsing and games to be using three cores quite often, and the fourth quite a bit. Android is quite multi-threaded, and if it also supports the Java Concurrency APIs it is very easy for software to also be multi-threaded.I also presume that the GeForce drivers and other Tegra SoC drivers utilise multi-threading as much as possible.
And Flash is being dropped on mobile devices in favour of HTML5. That's Adobe making that move. And not before time, it is a horrible technology.
Romulous - Thursday, November 10, 2011 - link
Meh. There may come a time when cores dont metter much.http://www.euclideon.com/ :)
alphadon - Thursday, November 10, 2011 - link
"Die size has almost doubled from 49mm^2 to somewhere in the 80mm^2 range"49^2 = 2401
80^2 = 6400
This should probably read:
"Die size has bloated to over 2.5 times the area of the prior generation leaving everyone wondering why NVIDIA is releasing this 40nm dinosaur. We would have expected a die shrink to keep the power and space requirements in line with the industry's competitors, but seeing all that real estate squandered on such an evolutionary product is downright shameful."
Lugaidster - Friday, November 11, 2011 - link
Did you even read the other posts? The other competitors have bigger dies and less cores! and also, the geometry didn't change between tegra 2 and this.I find it great that they were able to double the shader core count, increase core count from 2 to 5 (it's slower but not less complex, see the die picture) and increase frequency while still having a smaller die than the competition.
I think that given the constraints, this might turn out to be a good product. Obviously only time will tell if it actually performs, but who knows...
Lugaidster - Friday, November 11, 2011 - link
By the way, its 49 mm² not 49² mm. So its actually less than twice as big.psychobriggsy - Friday, November 11, 2011 - link
Do you seriously think the Tegra 3 die size is 8cm by 8cm?49mm^2 is the area, not the edge dimension. In effect the die size has gone from around 7mm x 7mm to 9mm x 9mm. I.e., your little finger nail to your index finger nail (your hands may vary).
psychobriggsy - Friday, November 11, 2011 - link
In addition the 28nm shrink of Tegra 3 (Tegra 3+) next year, if no extra features are added, will shrink the die from 80mm^2 to 40mm^2 (in an ideal world, let's say 50mm^2 worst case and shrinks aren't simple). And Tegra 4 will probably be around 80-100 mm^2 again.lightshapers - Friday, November 11, 2011 - link
This quad core architecture is still disappointing. Actually they implemented a 4th core with good reinforcement from marketing, presented to be a solution for low power consumption at low CPU load. My guess is actually all competitors can do this (cut clocks and power on all but one CPU and reduce cluster frequency) on the ARM dual cluster without the need to add an extra CPU ( I speak for Samsung and TI, as Qualcomm is designing their own). In addition to that, this 5th core is another non negligible additional gates that leak.Then, the action to synchronize L2 Cache by arm coherency port is fast, but 1MB is 1MB, which means probably few hundreds of us for lost reactivity when switching between cluster and this 5th core.
And at the end, it doesn't really solve all the problems of having 4 cores, as asymmetry in core load balance will ever happen. This solution may solve the low load case. But over the low-load watermark, the cluster is power-up, and we have 4 core consuming at least their leakage. This was reported as an issue on tegra2, I don't think it has changed (the 5th is in some way the proof), but here we have 2 additional cores...
For example, medium load requires 2 cores. 5th is off, but consumption is 4 times the consumption of one.
It would have been smarter to design a full speed additional core, so as to get a higher "low limit load" so as to stay on 1st core as longer as possible. With 500Mhz, it's difficult to say if you can manage all graphical interface + OS background on a 720p device...