NVIDIA Tegra Architecture

[ams]

Quad-core A57 variant would likely be targeted at the same price segment as the quad-core R3 Cortex A15 variant we see today: $299 tablets and $299 Chromebooks. An A57 core at 20nm should be a pretty potent performer for mobile devices (especially in multi-core benchmarks using all the cores), and time to market will probably be six months before the next gen Denver core is ready (first silicon for the first gen Denver core was only 9 months ago), so it is a reasonably good stop-gap solution. And of course the GPU perf. and GPU perf. per watt will be exceptionally good with the Maxwell GPU inside.

 

[Ailuros]

Meaning that current Denver cores in K1 are not potent enough for an Erista successor and they need to further pump it up or am I missing here something?

 

[ams]

The Denver core should be more than a match for an A57 core with respect to single threaded perf. and browser perf, but the quad-core A57 implementation should have higher multi-core benchmark scores in comparison. If Chinese OEM's demanded a 64-bit quad-core SoC, then NVIDIA would have no choice but to go with an A57 variant to fill that need at this point in time, due to the relatively large die size of the Denver CPU cores.

That said, it is inevitable that Tegra will at some point use a Denver CPU + Maxwell GPU, but the question is "when" (we will probably learn more about their plans at 2015 CES in January or at GTC 2015 in March).

 

[Ailuros]

The Denver core should be more than a match for an A57 core with respect to single threaded perf. and browser perf, but the quad-core A57 implementation should have higher multi-core benchmark scores in comparison.

That's nothing new for dumb multi-core synthetics with little to no realistic relevance like Antutu.

If Chinese OEM's demanded a quad-core SoC, then NVIDIA would have no choice but to go with an A57 variant to fill that need at this point in time, due to the relatively large die size of the Denver CPU cores.

So why didn't chinese OEMs line up in masses to fetch a 32bit K1 in the meantime? How many of them actually use as expensive SoCs for their devices?

4*A57+4*A53 isn't going to be cheap in hw either.

There's something not adding up here. You usually don't go through as long and expensive lengths to develop a custom core to only use it in some cases unless you're as big as QCOM f.e. and actually have a decent enough foot in the smartphone market.

A Grey sucessor ie a mainstream SoC with ARM cores and a cut back Maxwell ULP GPU variant would had been a completely different story. But practically two high end SoCs within the same year sound weird at best to me at the moment.

 

[ams]

Xiaomi, Acer, and HP are all using the quad-core TK1 variant in high volume products this year (over the course of a full year, this would translate to millions of TK1 SoC's shipped). And since there is no built-in cost associated with an integrated modem, the SoC pricing should be quite reasonable given the very high GPU and CPU performance.

Realistically, if there was no OEM demand for a quad-core Tegra variant, then NVIDIA wouldn't build it. So if they are building it, then there has to be demand coming from somewhere, and the most likely place is China.

Denver will make it's way into not just Tegra but also Tesla, Quadro, and GeForce at some point down the road. The purpose of Denver is to first and foremost extend the reach of ARM, but this can only be done one step at a time.

 

[Ailuros]

Xiaomi, Acer, and HP are all using the quad-core TK1 variant in high volume products this year (over the course of a full year, this would translate to millions of TK1 SoC's shipped). And since there is no built-in cost associated with an integrated modem, the SoC pricing should be quite reasonable given the very high GPU and CPU performance.

What's a high volume exactly compared to Rockchip's or Allwinner's volumes in China exactly? And why the dependence on any far east market all of the sudden? North America or Europe wouldn't all of the sudden appreciate their products?

Realistically, if there was no OEM demand for a quad-core Tegra variant, then NVIDIA wouldn't build it. So if they are building it, then there has to be demand coming from somewhere, and the most likely place is China.

Realistically I'd rather wait if for this fiscal year they've finally reached the break even point for Tegra. With two different SoC variants R&D expenses are obviously significantly higher.

Denver will make it's way into not just Tegra but also Tesla, Quadro, and GeForce at some point down the road. The purpose of Denver is to first and foremost extend the reach of ARM.

That was the original plan obviously. What will become of it in the longrun is a chapter of its own.

 

[ams]

The bulk of the R&D cost is the GPU which is now shared and leveraged with other lines of business.

 

[Ailuros]

The bulk of the R&D cost is the GPU which is now shared and leveraged with other lines of business.

2 SoCs from design to manufacture don't cost less or equal to 1.

 

[Xmas]

OTOH it's not prohibitively expensive, either. Nvidia produces a very small number of SoC variants relative to other companies.

 

[Nebuchadnezzar]

If Denver can't serve as a low power platform for smartphones they would be smart to go A53/A57 if they want to have any chance at all to come back into the phone market.

 

[Alexko]

If Denver can't serve as a low power platform for smartphones they would be smart to go A53/A57 if they want to have any chance at all to come back into the phone market.

I think it's pretty clear that they've given up on phones.

 

[ams]

If Denver can't serve as a low power platform for smartphones they would be smart to go A53/A57 if they want to have any chance at all to come back into the phone market.

Denver is expected to be more power efficient than a Cortex A15 (and presumably A57) core on the same fab process node, so I don't think that the custom core is the issue (and Apple has already proven that a "desktop-class" CPU core can work quite well in a smartphone). The issue in getting into smartphones is due in large part to the baseband modem tech. Qualcomm's WCDMA tech is literally required for use with legacy 3G networks used by some very large cellular service providers such as Verizon and Sprint. Since Qualcomm can effectively bundle their modem with their application processors, it makes it very difficult for other suppliers such as NVIDIA and Intel to get a foothold into the smartphone market. NVIDIA's expertise is graphics technology, and that is the biggest contribution that they can make to the ultra mobile space, whether through graphics IP and graphics core licensing or through use of their own very high performance application processors.

 

[Ailuros]

Denver is expected to be more power efficient than a Cortex A15 (and presumably A57) core on the same fab process node, so I don't think that the custom core is the issue (and Apple has already proven that a "desktop-class" CPU core can work quite well in a smartphone).

Whereby Cyclone even in A8 gets to 1.4GHz peak "only" while sustaining competitive performance in real time compared to other solutions. We don't know what Denver's efficiency looks like, but its peak frequency of 2.5GHz for K1 doesn't tell me that it's comparable in the above sense to Denver. Cyclone has in a very relative sense desktop class performance while sustaining ULP SoC class power consumption.

The issue in getting into smartphones is due in large part to the baseband modem tech. Qualcomm's WCDMA tech is literally required for use with legacy 3G networks used by some very large cellular service providers such as Verizon and Sprint. Since Qualcomm can effectively bundle their modem with their application processors, it makes it very difficult for other suppliers such as NVIDIA and Intel to get a foothold into the smartphone market. NVIDIA's expertise is graphics technology, and that is the biggest contribution that they can make to the ultra mobile space, whether through graphics IP and graphics core licensing or through use of their own very high performance application processors.

So how did Mediatek get to its success story in only few years and especially in China? Mediatek's core business is smartphones and they have had in recent years even a higher growth rate than QCOM. Their SoC top out where other's mainstream SoCs leave off and they have ultra low prices. That's a recepy that gets you endless design wins especially for China white box OEM deals which are extremely low cost conscious.

NV failed royally with GoForce years ago for smartphones and they also did with any Tegra generation so far. We could come up with even more excuses as to why they failed but it's not like they ever had any glorious success in any other ULP market so far either.

If they would had played their cards right from the start they could have taken the market by storm. But for that it takes a wee bit more than a bunch of nice sounding marketing rubbish like the GPU IP related stuff which will remain a complete joke until we see at least one serious IP licensing deal.

 

[ams]

Cyclone has in a very relative sense desktop class performance while sustaining ULP SoC class power consumption.

Denver is expected to have higher performance than Cyclone while sustaining ULP SoC class power consumption too.

So how did Mediatek get to its success story in only few years and especially in China?

Mediatek is the lowest cost and lowest margin SoC supplier, and their focus is on China, where the cellular service providers are not necessarily locked into Qualcomm's modem tech.

If they would had played their cards right from the start they could have taken the market by storm.

Not at all. There wasn't much that they could do to get around the issue of Qualcomm's modem bundling strategy and Qualcomm's tech being required in the North American market. Even Intel wasn't able to do anything about that.

 

[Ailuros]

Denver is expected to have higher performance than Cyclone while sustaining ULP SoC class power consumption too.

Remains to be seen as there's no substantial data even indicating yet any of the kind.

Mediatek is the lowest cost and lowest margin SoC supplier, and their focus is on China, where the cellular service providers are not necessarily locked into Qualcomm's modem tech.

I thought only a page ago China is/was amongst their target markets. Obviously MTK did something right which NV didn't.

Not at all. There wasn't much that they could do to get around the issue of Qualcomm's modem bundling strategy and Qualcomm's tech being required in the North American market. Even Intel wasn't able to do anything about that.

So now the focus suddenly changes back to North America again?

Intel failed repeatedly with smartphone SoCs because up to a point they were far too expensive and consumed too much power compared to competing solutions. Intel is selling more than well with tablets these days, but there are doubts that they reach their >40Mio units target and I'd love to see how their sales figures will look like without any subsidies in the future.

Both Intel and NVIDIA never bothered with any cheap low end solutions to address mass markets like China. Yes those are low margin but high volume; from that point and on an IHV can build up from humble budget solutions all the way up to high end. Taking the exact opposite route has obviously a higher risk.

It should NOT mean that modems are completely irrelevant, but in the given case not the real or sole scapegoat for any failure. For the very least NV didn't acquire Icera for nothing and if Tegra4i's lack of success is only the modem then questions arise as to why CDMA wasn't supported which is essential in the US.

 

[ams]

Remains to be seen as there's no substantial data even indicating yet any of the kind.

Oh please. The whole world knows that Tegra K1 Denver is the SoC launch platform for Android.L and the Nexus 9 tablet. So if it can fit inside a thin fanless 8.9" tablet, then it qualifies as having "ULP SoC class power consumption".

I thought only a page ago China is/was amongst their target markets. Obviously MTK did something right which NV didn't.

Mediatek has traditionally targeted the lower end of the Chinese market which is extremely cost conscious, which is totally different than what NVIDIA is doing. NVIDIA spends billions of dollars on GPU/CPU processor related research, and there is no way that they could do so on ultra slim margins.

So now the focus suddenly changes back to North America again?

The North American market is nearly as important as the Asian market when it comes to NVIDIA's overall revenues, and unlike Mediatek, much of NVIDIA's customer base has a significant presence in the North American market.

Intel failed repeatedly with smartphone SoCs because up to a point they were far too expensive and consumed too much power compared to competing solutions.

Completely false. Silvermont cores have very good power efficiency compared to Cortex A15 cores. And even with PowerVR GPU's and with contra revenue being included, Intel still cannot make a dent in the smartphone space due to the baseband story.

For the very least NV didn't acquire Icera for nothing

Obviously there is a future with VoLTE and where large cellular service providers move away from ancient legacy 3G networks that require Qualcomm tech, but that is still some way off.

FWIW, here is what was reported about Qualcomm:

"Qualcomm is the dominant global player in the mid and high-end, and uses its considerable intellectual property leverage to pursue significant licensing fees based on the wholesale price of 3G/4G mobile phones, leading authorities in China to charge Qualcomm with abusing its "dominant position in the market" and start anti-trust investigations in 2014. In July 2014, Qualcomm announced that it was having trouble enforcing royalties and license fees from Chinese manufacturers, impacting its sales expectations for the next quarter."

 

[Ailuros]

Oh please. The whole world knows that Tegra K1 Denver is the SoC launch platform for Android.L and the Nexus 9 tablet. So if it can fit inside a thin fanless 8.9" tablet, then it qualifies as having "ULP SoC class power consumption".

The whole world doesn't have a single independent measurement for Denver yet. Do you mind if I prefer to see real numbers first? You claimed higher performance for Denver against Cyclone (which I obviously take as a clock for clock basis) for which I said I prefer to wait. For something though that has ULP SoC class power consumption the excuse for using a 4+4 A57/53 config might get lost quickly again or simply belongs to another funky twist of the Erista story.

Mediatek has traditionally targeted the lower end of the Chinese market which is extremely cost conscious, which is totally different than what NVIDIA is doing.

One succeeded and the other failed.

The North American market is nearly as important as the Asian market when it comes to NVIDIA's overall revenues.

It's highly important to move goal posts to that degree in order to fit each excuse for any occassion.

Completely false. Silvermont cores have very good power efficiency compared to Cortex A15 cores. And even with PowerVR GPU's and with contra revenue being included, Intel still cannot make a dent in the smartphone space due to the baseband story.

There was a LONG line of SoCs before Moorefield/Medfield ever appeared.

Who told you that IMG GPU IP has the magic wand to save any wrong design decisions? Do you recall how long Intel chewed on the SGX535 while only raising its frequency?

 

[ams]

The whole world doesn't have a single independent measurement for Denver yet. Do you mind if I prefer to see real numbers first?

NVIDIA has already gone on record as saying that TK1 Denver will have improved battery life compared to TK1 Cortex A15, so that implies improved CPU power efficiency.

You claimed higher performance for Denver against Cyclone (which I obviously take as a clock for clock basis) for which I said I prefer to wait.

Not on a clock for clock basis, but on a performance basis within an ultra mobile power envelope. What matters in a power constrained environment is perf. per watt, not perf. per MHz.

 

[Ailuros]

NVIDIA has already gone on record as saying that TK1 Denver will have improved battery life compared to TK1 Cortex A15, so that implies improved CPU power efficiency.

NV isn't obviously an independent source in the given case.

* Is the only change in 64bit K1 just the two Denver cores instead of the 4+1 A15 config or have there been other possible changes too?

* Did they state by which degree battery life has improved?

* What are the final CPU + GPU frequencies for Denver in Nexus9 and what does it's battery life look like?

Not on a clock for clock basis, but on a performance basis within an ultra mobile power envelope. What matters in a power constrained environment is perf. per watt, not perf. per MHz.

On a performance basis it obviously will be faster than Cyclone but not necessarily more efficient clock for clock. In the meantime what still escapes that picture is the necessity of the big.LITTLE A5x config in Erista. If Denver is so efficient then why bother with standard ARM cores in the first place?

 

[ams]

If Denver is so efficient then why bother with standard ARM cores in the first place?

The only reason I can think of is that there is some demand for a high end quad-core or "octa-core" SoC from some large OEM(s) in China.