NVIDIA Tegra Architecture

[A1xLLcqAgt0qc2RyMz0y]

Nvidia Confirms HTC Nexus 9

From BSN originally:

While combing through Nvidia’s (NASDAQ: NVDA) legal documents posted on the company’s website, Bright Side of News*’ sources pointed to a very interesting sentence that many may have overlooked or simply never seen.

Nvidia’s document specifically states, “the HTC Nexus 9, expected in the third quarter of 2014, is also expected to use the Tegra K1.” So, based on what’s known that there will without a doubt be a Nexus 9 made by HTC with Nvidia’s Tegra K1, but we also know that it should be expected in the third quarter of this year.

http://www.brightsideofnews.com/2014/09/10/nvidia-confirms-htc-nexus-9

The link to the "Document in question: http://nvidianews.nvidia.com/imagelibrary/downloadmedia.ashx?MediaDetailsID=2996&SizeId=-1&SizeID=-1

Now just shows: Media does not exists. So it looks like Nvidia took it down.

This link: http://www.engadget.com/2014/09/10/nvidia-hints-at-htc-nexus-9/?ncid=rss_truncated

got a snapshot of it before it was removed.

 

[A1xLLcqAgt0qc2RyMz0y]

Maxwell on next Tegra

With the release yesterday of the Maxwell 2 GPUs it should result in major improvement in the next Tegra over the current Kepler GPU which by itself is no slouch.

My understanding is that the next Tegra will be on 20nm so that along with the 2x improvements of Maxwell over Kepler should result in a very powerful SOC in performance yet be power efficient.

And ideas/guesses on what the Maxwell in the next Tegra will look like?

 

[Ailuros]

As discussed in the past one NV slide suggested 50% more FLOPs/W from the GPU for Erista. The likeliest scenario in that regard is a 2 SMM ---> 256 SP scenario.

 

[xpea]

As discussed in the past one NV slide suggested 50% more FLOPs/W from the GPU for Erista. The likeliest scenario in that regard is a 2 SMM ---> 256 SP scenario.

agree on everything above and hopefully, on the CPU side, 4 Denver cores will be used. With this kind of setup on 20nm, I see no competition in the performance area. Big question is, can this beastly SoC will fit in phones or will it be restricted to tablets/cars/chromebooks ?

 

[Arun]

Big question is, can this beastly SoC will fit in phones or will it be restricted to tablets/cars/chromebooks ?

It's not a question of peak power consumption to fit in phones, you can always clock it lower, or even disable 1 out of 2 SMMs (with power gating or even permanently for a lower-end SKU). It's really more of a question of power/FPS, SoC cost, and being competitive with Qualcomm's entire platform (baseband, many-band LTE front-end, etc). It's also important to have good commercial relations with your potential customers.

BTW, it's interesting that there was no mention of ASTC for GM204 while Tegra K1 already supports ASTC LDR. I guess they understandably didn't want to pay the (large!) area cost...

 

[Ailuros]

agree on everything above and hopefully, on the CPU side, 4 Denver cores will be used. With this kind of setup on 20nm, I see no competition in the performance area. Big question is, can this beastly SoC will fit in phones or will it be restricted to tablets/cars/chromebooks ?

With an as phat IPC even 2 CPU cores are more than sufficient. Why not go back to way less die area if they can?

As for smart phones in addition to Arun's post above, I don't think they'll reconsider a Grey successor any time soon.

 

[Lazy8s]

While nVidia doesn't have a compelling hardware platform for phones, they seem to still be the only SoC designer targeting performance-focused modern (thin) tablets well. This should keep landing them many of the flagships.

 

[Silent_Buddha]

While nVidia doesn't have a compelling hardware platform for phones, they seem to still be the only SoC designer targeting performance-focused modern (thin) tablets well. This should keep landing them many of the flagships.

I wouldn't count Intel out. Broadwell-Y is pretty compelling for thin tablets and only requires aluminum for passive cooling. Although we'll have to wait for tablets to hit the wild to really test it out.

Maxwell should help out Nvidia there for the next Tegra, however. Hopefully, they'll be able to ditch the exotic magnesium used for cooling the relatively hot K1.

Regards,
SB

 

[Ailuros]

While nVidia doesn't have a compelling hardware platform for phones, they seem to still be the only SoC designer targeting performance-focused modern (thin) tablets well. This should keep landing them many of the flagships.

Quite likely; however without any serious volume anywhere it sounds strange at least how they plan to keep up the department like that in the longrun. If things would be as easy NV could had developed it's top dog GPU chips to professional markets only and call it a day.

I wouldn't count Intel out. Broadwell-Y is pretty compelling for thin tablets and only requires aluminum for passive cooling. Although we'll have to wait for tablets to hit the wild to really test it out.

Maxwell should help out Nvidia there for the next Tegra, however. Hopefully, they'll be able to ditch the exotic magnesium used for cooling the relatively hot K1.

Regards,
SB

No idea what their exact plans are, but if the CPU should remain a dual core Denver as in 64bit K1 and the probability of a dual SMM Maxwell (higher efficiency and the ability to power gate one of the two clusters) combined with 20SoC (whatever that can deliver apart from much higher density), they should have a pretty compact SoC that I would dare to guess could lie at under 100mm2 also.

On a sidenote if they wouldn't use as aggressive frequencies and/or aggressive DVFS settings especially for the GPU block, they wouldn't most likely need any exotic cooling measures either.

 

[Erinyes]

No idea what their exact plans are, but if the CPU should remain a dual core Denver as in 64bit K1 and the probability of a dual SMM Maxwell (higher efficiency and the ability to power gate one of the two clusters) combined with 20SoC (whatever that can deliver apart from much higher density), they should have a pretty compact SoC that I would dare to guess could lie at under 100mm2 also.

On a sidenote if they wouldn't use as aggressive frequencies and/or aggressive DVFS settings especially for the GPU block, they wouldn't most likely need any exotic cooling measures either.

No idea about the Denver version but my info is that the A57 version taped out quite a while ago (sometime in July) on 20SoC as expected. I'm guessing the launch will be at CES with devices shipping by late Q1 or Q2.

The big news is that apparently Nvidia has abandoned the penta core/companion core concept and adopted big.LITTLE this time around and the configuration is 4 A57 + 4 A53 cores. Technical reasons aside, I suspect marketing also had a role to play here as marketing it is an "Octa core" chip is advantageous. The other confirmed fact tha I've heard is that it has full hardware encode/decode support for 4k H.265 (not much of a surprise here).

On the GPU I agree with your speculation on 2 SMM's (effectively a GM108?). Another point that remains to be seen is whether they've gone with LPDDR4, which I suspect they have. Overall, seems pretty much like Snapdragon 810, minus the baseband of course, and it should have considerably higher graphics performance.

 

[Ailuros]

No idea about the Denver version but my info is that the A57 version taped out quite a while ago (sometime in July) on 20SoC as expected. I'm guessing the launch will be at CES with devices shipping by late Q1 or Q2.

The big news is that apparently Nvidia has abandoned the penta core/companion core concept and adopted big.LITTLE this time around and the configuration is 4 A57 + 4 A53 cores. Technical reasons aside, I suspect marketing also had a role to play here as marketing it is an "Octa core" chip is advantageous. The other confirmed fact tha I've heard is that it has full hardware encode/decode support for 4k H.265 (not much of a surprise here).

Let me get this str8: Erista will have standard ARM cores?

 

[Erinyes]

Let me get this str8: Erista will have standard ARM cores?

Indeed. There will probably be two versions again like TK1 has but I have not heard anything about the Denver version yet.

 

[Ailuros]

Indeed. There will probably be two versions again like TK1 has but I have not heard anything about the Denver version yet.

What for? (honest question). For K1 it is my understanding that they had a second SoC variant with Denver cores in order to have 64bit CPUs in time; for Erista that field seems to be covered already by standard A5x ARM cores.

 

[xpea]

What for? (honest question). For K1 it is my understanding that they had a second SoC variant with Denver cores in order to have 64bit CPUs in time; for Erista that field seems to be covered already by standard A5x ARM cores.

I think it all depends on the performance delta between A57 and Denver. Plus the power consumption of course...

 

[Erinyes]

What for? (honest question). For K1 it is my understanding that they had a second SoC variant with Denver cores in order to have 64bit CPUs in time; for Erista that field seems to be covered already by standard A5x ARM cores.

Good question..and I'm afraid I dont have a definitive answer. Note that regarding your point about having 64 bit in time for K1..Samsung has actually beaten them to shipping a 64 bit SoC in a consumer product (with the 64 bit disabled though).

I suspect there could be a number of reasons though..timing for one. If like K1, the denver part follows by ~6 months, they may have been forced to go for a A57 part just to match the release window of Qualcomm/Samsung. Possible they're also working on a refresh of the Denver architecture which may be even more of a delay. Another reason could be marketing related ("Octa core" A57/A53 parts sound better to the average consumer). My info could also be plain wrong for all you know..but I have no reason to doubt the source.

 

[Ailuros]

It's not the first time I read about standard ARM cores in Erista, meaning it's hardly a coincidence. QCOM for instance obviously doesn't have a custom 64bit CPU core ready yet. With NV having Denver and at the same time using standard ARM cores it sounds quite awkward.

 

[Alexko]

Perhaps Denver simply isn't competitive with A57.

 

[ams]

Perhaps Denver simply isn't competitive with A57.

Heh, I don't think it has anything to do with that. A Denver core clearly should have significantly higher single threaded performance and significantly better power efficiency in comparison to an A57 core, especially when fabricated on the same fab process node.

There are certain high volume markets such as China where OEM's find it desireable to offer quad-core and "octa"-core offerings to the public, so a quad-core or "octa"-core Cortex A series powered SoC may make sense there. And perhaps the Cortex A series variant can come to market more quickly and at a lower cost too. Anyway, having two variants of the Tegra SoC that are presumably pin-compatible may be a sign that NVIDIA is spreading their risk and not putting all their eggs in one basket.

Here is what NVIDIA's Steve Scott has said about this subject in the past:

"Our Denver project is really aimed at putting out a high-performance ARMv8 processor. Our Denver 64-bit ARM core will be higher performance than anything you can buy from ARM Holdings. That core is going to show up in Tegra, but it won't show up in all of the Tegra processors. We will still have Tegra processors that use stock ARM cores as well, like we use Cortex-A9 cores today, but Denver will show up in the high end. As an architecture licensee, the thing to remember is that you can tweak an ARM core to change its performance, but you can't change the architecture one lick."

 

[mboeller]

Perhaps Denver simply isn't competitive with A57.

Heh, I don't think it has anything to do with that. A Denver core clearly should have significantly higher single threaded performance and significantly better power efficiency in comparison to an A57 core, especially when fabricated on the same fab process node.

Shouldn't we just wait a few more weeks before we draw any conclusions? The Nexus9 with the Denver-cores is said to be announced or even available at the end of October.

so please hold the horses

 

[Ailuros]

If Denver is hypothetically "high end" then what are 4*A57 + 4*A53 supposed to be? Mainstream or low end? Even if (which is frankly complete BS) why would any IHV combine a =/<mainstream CPU with an as high end GPU like the Maxwell grandchild Erista is going to contain?

If far east markets like China want 8-core CPUs for marketing reasons they also want incidentially ultra cheap SoCs preferable with as low as possible power consumption. Something like a Mediatek MT6592 for instance.