Denver is not dead, he's just sleeping! : Tegra X1 uses A57

[Genx87]

I'm not sure why they'd go to a smaller, weaker CPU core though. Driver problems? Thermal?

Anandtechs article mentions time to market. Sounds like Nvidia wants a part available for the next round of phones and or car devices. I think it is a stretch to claim Denver is dead. Certainly in this product cycle Nvidia deemed it more beneficial to use an A57. But we dont know what the next product will use.

 

[NTMBK]

Doesnt change the fact that their gross margins are positive.

...you know that doesn't make it a success, right?

 

[Vesku]

Denver has been a shaky project for Nvidia, one of the predictions Charlie of SemiAccurate actually had mostly correct. K1 having both an A15 and Denver version backs this up. On the bright side they seem to be quite competitive in the SoC GPU department.

 

[ShintaiDK]

The GPU performance/watt vs competition is quite amazing. Its like 3x of the A8X

 

[NTMBK]

The GPU performance/watt vs competition is quite amazing. Its like 3x of the A8X

I wouldn't go that far, given that the X1 is rated for up to 10W:

Two and a half years ago, the Elemental demo of Unreal Engine used to consume over 300W of power. A year ago, it consumed 100W when demoed on the Xbox One. Today, it's at 10W with the new Tegra X1.

Given that the Air 2 has better battery life than its predecessor, I'm fairly sure the A8X is drawing a lot less than 10W.

But it looks pretty impressive, regardless. Hoping it shows up in some Android TV devices, should be good for gaming.

 

[TreVader]

What a surprise! Crazy, insanely high benchmarks with no data to back them up? From nvidia? Double Apple performance?



Yawn.



We heard this all before with K1. How many of you own a Nexus 9?


I rest my case.

 

[Shivansps]

But it looks pretty impressive, regardless. Hoping it shows up in some Android TV devices, should be good for gaming.

but thats the problem, what game that can run on ARM needs that much gpu power? what a waste.

 

[monstercameron]

but thats the problem, what game that can run on ARM needs that much gpu power? what a waste.

build it and they will come.

 

[Dribble]

Denver is dead: Tegra X1 uses A57

Looking for a job at fudzilla?

Sensationalist title aside looks like updating both Maxwell, and Denver at the same time while moving to a new node was too much for Nvidia.

 

[witeken]

The GPU performance/watt vs competition is quite amazing. Its like 3x of the A8X

Can't we just wait with such extraordinary claims until there's an independent review out there that measured performance and power? I've seen the Maxwell review, but I don't trust Nvidia's marketing spin at all.

 

[krumme]

Looking for a job at fudzilla?

Sensationalist title aside looks like updating both Maxwell, and Denver at the same time while moving to a new node was too much for Nvidia.

To much what?
They ported everything else including their custom interface and isp. Except the cpu. If that part is not worth porting but the rest is it looks damn bad and its a shame because its very innovative tech. Whatever the reason. Tick tock ttm or raw cost. Its just as bad or dead as it gets.

But innovating like that is a gamble; little chance of success but if they had won the benefits would have been enormous. I think they got pretty close. Well worth the try and imo shows the decision to go for the win was right. But i wouldnt bet they can resurrect this dead dude.

 

[Shivansps]

build it and they will come.

on android? forget it, no one gona build a game for the 1% of android users. Take a look at how relevant where the older Tegras, the exact same thing happen, by the time it becomes relevant, there are much better options.

Im just not interested in a lot of IGP power if not an x86, and im not the only one here.

 

[Dribble]

To much what?
They ported everything else including their custom interface and isp. Except the cpu. If that part is not worth porting but the rest is it looks damn bad and its a shame because its very innovative tech. Whatever the reason. Tick tock ttm or raw cost. Its just as bad or dead as it gets.

But innovating like that is a gamble; little chance of success but if they had won the benefits would have been enormous. I think they got pretty close. Well worth the try and imo shows the decision to go for the win was right. But i wouldnt bet they can resurrect this dead dude.

I take it they would have needed Denver mk 2 with updated everything, more cores, different memory interface, etc. I'm no HW expert but I bet doing all that and getting it to work on a new node is non-trivial. It's not like plan B (use arm cores with nvidia interconnect) is exactly going to hurt them - it is still (allegedly!) faster then anything else out there.

 

[Arachnotronic]

Looking for a job at fudzilla?

Sensationalist title aside looks like updating both Maxwell, and Denver at the same time while moving to a new node was too much for Nvidia.

They ported Maxwell, built and entirely new LPDDR4 interface, ported/improved all of the other IP blocks, and you're telling me they couldn't port a CPU core in time as well? Also, the A57 they did was probably a custom layout, not an off-the-shelf physical IP from ARM.

EDIT: BTW, love how the only thing people are focused on is whether NVIDIA did a custom CPU or not. The rest of the SoC is crazy complex and also looks very compelling. There is nothing shameful about using ARM CPU cores which are quite good. ARM has been designing CPUs far longer than NVIDIA has. NVIDIA should focus on differentiating in areas where off-the-shelf IP is not available or not good enough. A57 is good enough for this class of SoC and has room to run on future process nodes.

 

[Headfoot]

Porting a CPU takes time. This is not negotiable. There is no hand waving it away. It takes time. Likely, the project managers sat down and concluded the product would be least affected by swapping CPUs for something that would reduce time to market. What's the alternative -- swapping a generic ARM GPU? They probably want to do substantive revisions to Denver in addition to shrinking it, adding further time

They couldn't port it in time as is apparent by the fact that nvidia said "We chose ARM a57 for time to market reasons" and "Denver is still in our roadmaps." Tegra has gotten burned by time to market woes before so they seem to be wising up and releasing when it needs to be released leaving some parts on the cutting room floor if necessary

 

[tviceman]

They ported Maxwell, built and entirely new LPDDR4 interface, ported/improved all of the other IP blocks, and you're telling me they couldn't port a CPU core in time as well? Also, the A57 they did was probably a custom layout, not an off-the-shelf physical IP from ARM.

You have to keep in mind Nvidia only has so many engineers and R&D money for work on architectures. Nvidia inserted Erista in between Parker and Logan because they knew TSMC wasn't going to have finfets ready. They probably brought the Denver core up in release with schedule with Logan (aka Tegra 5) to A) get the first Google x64 tablet onto the market and B) learn where and what needs improvement on the Denver core for it's next iteration. Given all that, it's very plausible that they are still heavily working on tweaking / improving Denver and choose to spend their resources entirely on Denver's finfet design rather than porting to 20nm then porting it to finfets. I believe that Parker will have a Denver variant, if not be available only with Denver cores.

Anyways, Erista looks great. Here's hoping it's in more tablets stateside than TK1 was in. Nexus 9 would have been a buy for me if the storage options weren't so ridiculously limited.

 

[krumme]

Agree. Denver is dead but the rest is worth pursuing. Make Maxwell the standard. Use custom arm cores and use their software expertice to get in eg automobile.
The profit is not in the hardware anyway. Nv brand and software experience is worth far more than another cpu core.

 

[krumme]

You have to keep in mind Nvidia only has so many engineers and R&D money for work on architectures. Nvidia inserted Erista in between Parker and Logan because they knew TSMC wasn't going to have finfets ready. They probably brought the Denver core up in release with schedule with Logan (aka Tegra 5) to A) get the first Google x64 tablet onto the market and B) learn where and what needs improvement on the Denver core for it's next iteration. Given all that, it's very plausible that they are still heavily working on tweaking / improving Denver and choose to spend their resources entirely on Denver's finfet design rather than porting to 20nm then porting it to finfets. I believe that Parker will have a Denver variant, if not be available only with Denver cores.

Anyways, Erista looks great. Here's hoping it's in more tablets stateside than TK1 was in. Nexus 9 would have been a buy for me if the storage options weren't so ridiculously limited.

Serious. A cpu not worth porting in a market where ttm is crucial?
Yes its shows its difficult and expensive but thats the point. Its not worth it compared to other priorities.
At the time tsmc finfet is ready for denver2 we are closer to apple a10 than 9.

I am pretty confident apple could make the same demonstration on a9 now as nv does on the current a57/maxwell setup. On 16nm.
Nv is late as hell. Actual products if any is what 6 month+ from market? The decision not to use denver but a57 must have delayed it. The timeframe shows unfortunately OP is right.

 

[Headfoot]

. The timeframe shows unfortunately OP is right.

It shows nothing of the sort. That's one speculative conclusion on the facts. Another equally plausible speculative conclusion is that nVidia had to come up with a 20nm product quickly due to TSMC's ever-changing roadmaps for 20, 16 and 16FF+. I imagine 16nm FF Denver is under development as we speak and couldn't be finagled to work on 20nm without more effort than its worth for a supposedly short-lived node. Now whether it is ultimately a successful chip that's faster than the competitors is a separate question: but I have no doubt nVidia is still going to push Denver.

It gives them a second product launch with new bullet points vs the competition:
A57 Maxwell = "Better GPU!"
Denver2 Maxwell = "Better GPU AND better CPU!" (provided it's actually faster of course)

And they already did it once with K1-32 and K1.

I very much doubt Denver is going to be thrown aside after so many dollars invested into it. I'd call it a reasonable hedge.

 

[krumme]

It shows nothing of the sort. That's one speculative conclusion on the facts. Another equally plausible speculative conclusion is that nVidia had to come up with a 20nm product quickly due to TSMC's ever-changing roadmaps for 20, 16 and 16FF+. I imagine 16nm FF Denver is under development as we speak and couldn't be finagled to work on 20nm without more effort than its worth for a supposedly short-lived node. Now whether it is ultimately a successful chip that's faster than the competitors is a separate question: but I have no doubt nVidia is still going to push Denver.

It gives them a second product launch with new bullet points vs the competition:
A57 Maxwell = "Better GPU!"
Denver2 Maxwell = "Better GPU AND better CPU!" (provided it's actually faster of course)

And they already did it once with K1-32 and K1.

I very much doubt Denver is going to be thrown aside after so many dollars invested into it. I'd call it a reasonable hedge.

Its not like they skipped tsmc 20nm all together because of bad excuses neither top management or shareholders care to hear.
They ported ALL sans cpu and even added a custom a57 core. Tell me where the 20nm desktop and laptop mobile gpu is?
Yes its all speculation and there seems to be lots of potential in denver. But there have been plenty time to port it.
Now it can be we see 16 finfet denver but anyway this is at best a bad sign for us. Most likely this is nv usual marketing machine at full speed and denver is as dead as something like that can be.

 

[Lepton87]

Isn't quad A57 actually faster then dual core denver, at least its current revision? They would have to either double the amount of cores or improve IPC/ increase clock-speed to beat BIG.Little A57/53. I think A57 is actually better than denver for a mobile SOC.

 

[witeken]

Quadcores are still more marketing than useful. Don't get me started about Octacores:

 

[MisterLilBig]

I also believe that not using Denver means something bad about Denver. Think of the product, the X1, as it was shown, wouldn't this had been another chance to increase developer mindshare for their own custom cores? I think so.

And, this was shown as a platform for auto-computing, mostly, why didn't they just go with 8 A57 cores or 8 A53 cores? Seems like they went ahead with the most common set-up to see if they could attract some of the usual vendors along the way.

 

[krumme]

Isn't quad A57 actually faster then dual core denver, at least its current revision? They would have to either double the amount of cores or improve IPC/ increase clock-speed to beat BIG.Little A57/53. I think A57 is actually better than denver for a mobile SOC.

Its my impression denver uses to much power for phone. And what is the market for nv selling cpu for tablets when china dumps cheap cores on the market and Intel in reality even pays for each sold cpu and where Apple use their own cpu and SS prefers the same? Its crazy market and not worth it. Going for a niche like auto seems fine and can potentially turn a good profit.

 

[NTMBK]

Looking for a job at fudzilla?

Sensationalist title aside looks like updating both Maxwell, and Denver at the same time while moving to a new node was too much for Nvidia.

Yeah yeah, I changed it to something else