Print 17 comment(s) - last by name99.. on Jan 7 at 8:22 PM

NVIDIA looks to strengthen its position in the mobile sectors with Tegra 4

With smartphones and tablets taking over the computing world and putting traditional PCs and notebooks on the back burner, a number of companies are jockeying for position to deliver the highest performing SOCs on the market. NVIDIA has been in this game for a while with its Tegra line of processors, and its most recent Tegra 3 has scored a slew of design wins over the past year.
NVIDIA is looking to build upon that success with its next generation Tegra 4. While we're sure that NVIDIA was looking to surprise everyone at CES, most of the details on the new chip leaked in mid-December. The Cortex-A15-based Tegra 4 is built on a 28nm process, continues the 4+1 design (quad-core + companion core), and features 72 NVIDIA GeForce GPU cores.

NVIDIA says that the move to 28nm help the Tegra 4 consumer 45 percent less power than its predecessor and would allow mainstream phones to deliver 14 hours of continues HD video content.
However, Tegra 4 doesn't have integrated LTE onboard. Instead, NVIDIA is hyping up its optional Icera i500 processor for LTE functionality. Although the move to 28nm is likely to improve battery life across the board, not having LTE integrated on-chip isn't going to do it any favors either.
We’re likely to see a number of Tegra 4-based products at CES, and we’ll be sure to keep you informed of the ones that really catch our eye.

Source: NVIDIA

Comments     Threshold

This article is over a month old, voting and posting comments is disabled

What no LTE on die means...
By nafhan on 1/7/2013 11:15:02 AM , Rating: 2
No LTE on-die means this is absolutely not a cell phone chip; it's for tablets/laptops. 4 X A15 is already a pretty good indicator of that (4 A15's is a LOT of heat), but a separate chip for LTE pretty much verifies that to be the case.

By aegisofrime on 1/7/2013 11:37:13 AM , Rating: 2
When I see the power consumption figures for the Exynos 5250 it's pretty scary... 3W usage on the Nexus 10! But then again Nvidia claims that the Tegra 4 will consume even less power than the Tegra 3. I'm looking forward to an indepth Anandtech article on the Tegra 4. I'm really interested in the die size of 28nm A15 vs 32nm A9...

RE: What no LTE on die means...
By theapparition on 1/7/2013 1:24:58 PM , Rating: 3
No LTE on-die means this is absolutely not a cell phone chip

Not entirely the case. They could always add a LTE chip to the design.

There's actually a lot of benefits as it enables a designer to pick the best performing LTE silicon available, in the proper frequency bands, etc. and stay processor agnostic.

For example, Qualcomm's LTE embedded design isn't fully compatible with all LTE implementations around the globe. In those instances, you have to use something else with 3rd party LTE design.

Drawbacks are extra space and generally higher power requirements. Engineering is always an exercise in compromise.

RE: What no LTE on die means...
By chmilz on 1/7/2013 2:19:19 PM , Rating: 2
That software programmable communication chip is probably bigger news than Tegra 4, we just don't know it yet.

RE: What no LTE on die means...
By nafhan on 1/7/2013 5:21:42 PM , Rating: 2
True. I don't have any inside info or anything. :) I do think that's something that will make it's usage less likely. I'm very much looking forward to third party reviews.

Surface 2 with Tegra 4 FTW
By Pirks on 1/7/2013 1:48:20 AM , Rating: 1
Can't wait to see how Vendetta Online looks on it. Not that that it looks bad on current Surface, but the space is the limit.

RE: Surface 2 with Tegra 4 FTW
By nikon133 on 1/7/2013 4:01:24 AM , Rating: 2
I'm really pleased developers did RT-friendly version. How does it look and play on Surface, compared to desktop version(s) - if you had a chance to try more than one?

RE: Surface 2 with Tegra 4 FTW
By Pirks on 1/7/2013 12:22:54 PM , Rating: 2
It looks really great for a mobile MMO space sim, same as on any other Android Tegra 3 device. Frame rate is smooth, all in all very enjoyable experience for a tablet game.

By theapparition on 1/7/2013 10:36:34 AM , Rating: 5
I disagree with your comments.

The Tegra 3 was actually more powerful CPU wise and competitive GPU wise with Apple's A5. So using the term "smash" was a bit incorrect. The A5X brought superior GPU capability by adding the SGX543MP4 quad core GPU. But CPU wise, it was basically still OMAP or Tegra 3.

The A6 upped the ante, certainly. A custom core based on the ARM instruction set rather than the reference A9/A15 design has it's benefits (and a few drawbacks). They actually downgraded the GPU to the SGX543MP3 but with a higher clock.

The A6 is the benchmark for newer SoCs, and kudos to Apple for that and setting a benchmark. But there's certainly a lot of upcoming competition. OMAP 5, Exynos 5, and Qualcomm's newest, along with the Tegra 4.

As to your question about GeForce cores, each Tegra 4 will come with 72. I have no knowledge of them custom making dies for customer applications. However, depending on yields, it may be possible that there will be chip fallout and binning will occur. In that case, there may be different binned versions. From past NVidia history, this isn't a difficult chip for TSMC to make and yields should be quite good.

By Wolfpup on 1/7/2013 11:23:06 AM , Rating: 3
Apple does NOT have the fastest mobile SOCs around. They're good, but...

The "smashed Tegra 3" thing is bogus and was addressed.

Apple's new Swift CPU cores are faster than A9, but nowhere near as good as A15...and this has FOUR A15s versus two Swift cores. That ain't "smashing" anything.

Don't know how the GPU cores will compare, but at least by shear number this is way past Tegra 3, so seems likely it's the best right now.

By nafhan on 1/7/2013 11:35:26 AM , Rating: 2
Apple took a track similar to what Qualcomm did in regards to the A15: ARMv7 compatible CPU that has a MUCH better power consumption profile than a "real" A15 at the expense of some speed. I don't think we've seen an A15 design, yet, that would make sense in a smartphone (including Tegra 4). So, they're definitely doing the right thing.

I'm really curious to see what power numbers for the Tegra 4 look like. The fact that they're going on and on about performance means it probably won't have great power numbers, IMO.

By zephyrprime on 1/7/2013 12:25:26 PM , Rating: 2
Well, the article does say that it uses 45% less power.

By nafhan on 1/7/2013 5:19:18 PM , Rating: 2
I dunno... I'm still holding out for real numbers. HD video might be handled by "5th" core + GPU or something.

By nafhan on 1/7/2013 11:29:19 AM , Rating: 2
I think we're making the switch from fixed function shaders. So, based on my experience when this happened on the desktop 2 Tegra 4 unified shaders will equal 1 Tegra 3 fixed function shader pipeline. So, this probably has about 3X the GPU power of the Tegra 3 in shader limited situations (72/2 = 12*3), which would put it in the same ballpark as the iPhone 5. That also matches up with what Nvidia did on the Tegra 3: same ballpark as the iPhone 4S's chip.

Nvidia is claiming 6X the GPU power of the Tegra 3, but I'm thinking that's probably best case scenario and/or they're running it at much higher clockspeeds than Tegra 3, or maybe they've dramatically increased memory bandwidth. I know older Tegra designs were somewhat limited in that regard. It'll be interesting to actually see 3rd party benches.

By TakinYourPoints on 1/7/2013 12:11:05 PM , Rating: 2
Aside from Apple there's also Qualcomm that NVIDIA is playing catch-up with. I'm looking forward to benchmarks by Anandtech, but given the pattern of the last few years at best I expect it to just barely catch up with what there is now, only to be leapfrogged shortly by Apple and Qualcomm.

We'll see.

By name99 on 1/7/2013 8:22:30 PM , Rating: 2
Up until the A6, they literally did very little work, and yet smashed Tegra3 in pretty much every way.

This is not a helpful way to characterize Apple's behavior.
Apple started off with extremely off-the-shelf parts, the innovation was in the software (and perhaps in the choice of additional parts added, the breadth of sensors, the use of a high-quality, for the time, h.264 decoder).
They moved on to asking for parts that had higher performance (most obvious in the GPU arena).
They moved on to a custom SOC (made with mostly 3rd party cells, but with some unorthodox choices, like higher memory bandwidth).
And then we got a custom CPU.

You see this as four years of doing nothing, then a custom CPU. I see it as four years of learning on the job, one step after another.

As for the "my 4 A15s can beat up your 2 Swifts", grow up.
There is no perfect CPU design. You can optimize for power. You can optimize for computational performance. You can optimize for memory performance. There's a reason ARM AND Intel AND IBM are all three successful in the CPU business.
The trick, however, is to know what your market requires, otherwise you produce a white elephant like the Pentium4.
If you're going to rant about how wonderful 4 A15 cores are, do so by
(a) providing numbers (power and memory numbers) and
(b) providing realistic CURRENT-DAY scenarios in which four cores are useful.

I'm all for competition in this space. A7, Tegra4, next-gen Atom --- bring them all on. But let's try to maintain a level of discourse higher than "15>7 therefore A15 is better than A7" or "two cores bad, four cores good".

Of particular interest here is the memory behavior of Tegra 4. As I have said many times before, an obvious lesson to me when I worked at Apple and we were using PPC was that memory performance really matters, and PPC's superior core was irrelevant in the face of Intel's substantially superior memory performance. Apple's constant push to ramp up the memory performance of iOS devices shows that they still remember this lesson.
It would not surprise me if a future path for Apple consists of a lot more apparently unsexy stuff going on in the memory uncore, and a lot less concern with pumping up the core. So while Tegra may look cool in terms of better branch predictors, a more aggressive superscalar implementation, more rename registers, etc, Apple will be putting its design energy into a smarter memory controller and better PoP packaging, to bring the DRAM closer to the core and allow it to run faster, while still not burning too much power in the memory bus.

The details of the core are sexy, I know --- I love that stuff. The details of the memory system seem so damn boring, from coherency protocols to memory controller state machines. But we are at the stage (and about fifteen years of experience has shown) that a sexy core hooked up to a sub-par memory system is usually very disappointing in real-world code, as opposed to micro benchmarks.

"We shipped it on Saturday. Then on Sunday, we rested." -- Steve Jobs on the iPad launch
Related Articles

Copyright 2016 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki