It's possible that they'll use the Denver part for repetitive optimized tasks and A57 for the other. I doubt this is even gonna be in a phone/tablet, they've got rid of the power saving cores. Looks like this has been designed ground up for automotive applications.
Why on earth would they use the A57's at this point? when A72 i out and, well better, and they have their own core, if youre going custom, go fully custom, if not, use the newest and high performing Core.
The A57 is used *with* the Denver cores, so they are probably used as companion cores for specific threads or operations. On top of that, Nvidia is most likely building a custom version of Android or custom kernel to run this hardware. With full control over the hardware and software layers, they can leverage the strengths of the A57.
Also, we can speculate that the A72 was either too new for Nvidia when they started this project or the A57 is better suited for their use case.
Remember, this is for automotive use, not mobile devices.
I hoped they would do this and run the new code on the ARM core (hoped A72 but , oh well) while also translating it for Denver so the next time it can run on Denver. Remains to be seen how they do power management,might have been easier to just give Denver a short clock boost when dealing with new code. Then again ,they don't get any wins so it matters less.
The SOCs used in this board could just meant for that board and not used anywhere else, like tablets. So knowing that those SOCs will be used in self driving cars, and considering the possibility Nvidia to also offer the software for that usage, then probably for that specific software and those specific tasks, Denver cores are better than A57s.
If it was just for cars they would have detailed it today but they seem to have delayed the reveal for MWC. They can't really afford to make a chip just for auto for now anyway, they need to try and get some extra revenue from elsewhere.
No they don't have to detail it today or even tomorrow. Those who are going to care about this are those in big offices in car companies. Those are the people who will be informed about it. Not the end user. The end user will just learn from commercials that this car is a magical car and nothing more. That customer is a car driver not a computer user.
Tegra PX targeted at both the Console and Automotive sector would make sense. It would need sufficient performance to allow porting of current Console games and, very important, get Game designers interested in porting. So maybe 512 CUDA cores, 128 bit LPDDR4 bus (up to 4.266 Gbps?) with color compression. 8GByte (maybe two mobile 4GByte stacks in the package?)
Yeah, whilst I think it's 384 @ 30% higher clocks, I would expect it to be LPDDR4 and if it's 1 TFLOPS like I suspect then you'd want the 56 GB/s a 128-bit LPDDR4 bus would give.
Since PX2 is 8tflops, and its a dual soc, would the Single tegra SOC be 4tflops, or at least around there, because if so, then that would make it way ahead of today's current gen consoles.
One thing I think it would be safe to assume is that Denver has been worked on in the 18+ months since Denver TK1, and that the core used in Drive PX 2 is not the same iteration as that used in TK1. It seems useless to speculate whether it has the same performance characteristics of the previous version or not.
But even if it did, other things to consider are that Drive PX 2 is a compute-oriented setup, and Pascal supports unified memory. The Drive PX 2 system is designed for the particular workload of machine learning for self-driving cars. The type of stuff going to be run on these machines is well known, so they should be able to much more finely tune what gets executed by what, I am guessing?
Finally, unlike last year with the Drive PX where NVIDIA said Drive PX was powered by two Tegra X1's, NVIDIA didn't announce an SOC name being used with the Drive PX 2. That seems like an indication that this particular SOC configuration isn't planned to be marketed outside of Drive PX 2 or similar in-house systems.
"this new SoC (Tegra P1?) has four Cortex A57s and two Denver CPUs. As of now it isn’t clear whether this is the same iteration of the Denver architecture that we saw in the Tegra K1. However, regardless of what architecture it is we’re still looking at a CPU architecture that is at least partially an ARM in-order core with a wide, out of order VLIW core that relies on dynamic code optimization to translate ARM instructions into the VLIW core ISA."
The Cortex-A57 CPU is an out-of-order CPU architecture, so the phrase "an ARM in-order core with a wide, out of order VLIW core" seems out-of-place and incorrect.
Nvidia seems to always be behind one way or another. Denver is a 2014 core, and A57 is a first half of 2015 core. This will will arrive in the second half of 2016 most likely, and will have to compete against Kryo and A72.
I blame this on Nvidia's lazy execution. They'd rather keep being lazy and use an old core, instead of running a tighter ship over there and launch a more contemporary CPU like everyone else.
Denver is a project name, and NVIDIA doesn't seem to use it the way you are using it here by putting it in the same category as the term "A57" and assigning it a particular date (2014). I think it's safe to assume that the 2016 Denver CPU will not be the same core as the 2014 Denver CPU. Besides, speculation on Parker based on Drive PX 2 is dubious being that NVIDIA didn't mention any SOC name when talking about Drive PX 2. Drive PX 2 has particular demands which are different from the demands of a tablet or smartphone device. Even if the Denver used in Drive PX 2 is the 2014 Denver, there is no good reason to assume that NVIDIA's next tablet/console SOC, assuming there is one, will use the same core or resemble Drive PX 2 in any way.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
23 Comments
Back to Article
hans_ober - Tuesday, January 5, 2016 - link
It's possible that they'll use the Denver part for repetitive optimized tasks and A57 for the other. I doubt this is even gonna be in a phone/tablet, they've got rid of the power saving cores. Looks like this has been designed ground up for automotive applications.Mobile-Dom - Tuesday, January 5, 2016 - link
Why on earth would they use the A57's at this point? when A72 i out and, well better, and they have their own core, if youre going custom, go fully custom, if not, use the newest and high performing Core.PaulJeff - Tuesday, January 5, 2016 - link
The A57 is used *with* the Denver cores, so they are probably used as companion cores for specific threads or operations. On top of that, Nvidia is most likely building a custom version of Android or custom kernel to run this hardware. With full control over the hardware and software layers, they can leverage the strengths of the A57.Also, we can speculate that the A72 was either too new for Nvidia when they started this project or the A57 is better suited for their use case.
Remember, this is for automotive use, not mobile devices.
jasonelmore - Tuesday, January 5, 2016 - link
yeah i'm not a fan of denver's code morphing architecture.. I wish they would ditch the concept all togather.jjj - Tuesday, January 5, 2016 - link
I hoped they would do this and run the new code on the ARM core (hoped A72 but , oh well) while also translating it for Denver so the next time it can run on Denver. Remains to be seen how they do power management,might have been easier to just give Denver a short clock boost when dealing with new code.Then again ,they don't get any wins so it matters less.
yannigr2 - Tuesday, January 5, 2016 - link
The SOCs used in this board could just meant for that board and not used anywhere else, like tablets. So knowing that those SOCs will be used in self driving cars, and considering the possibility Nvidia to also offer the software for that usage, then probably for that specific software and those specific tasks, Denver cores are better than A57s.jjj - Tuesday, January 5, 2016 - link
If it was just for cars they would have detailed it today but they seem to have delayed the reveal for MWC. They can't really afford to make a chip just for auto for now anyway, they need to try and get some extra revenue from elsewhere.yannigr2 - Tuesday, January 5, 2016 - link
No they don't have to detail it today or even tomorrow. Those who are going to care about this are those in big offices in car companies. Those are the people who will be informed about it. Not the end user. The end user will just learn from commercials that this car is a magical car and nothing more. That customer is a car driver not a computer user.Hans de Vries - Tuesday, January 5, 2016 - link
Tegra PX targeted at both the Console and Automotive sector would make sense. It would need sufficient performance to allow porting of current Console games and, very important, get Game designers interested in porting. So maybe 512 CUDA cores, 128 bit LPDDR4 bus (up to 4.266 Gbps?) with color compression. 8GByte (maybe two mobile 4GByte stacks in the package?)psychobriggsy - Tuesday, January 5, 2016 - link
Yeah, whilst I think it's 384 @ 30% higher clocks, I would expect it to be LPDDR4 and if it's 1 TFLOPS like I suspect then you'd want the 56 GB/s a 128-bit LPDDR4 bus would give.darkich - Tuesday, January 5, 2016 - link
Alright AT..where the heck is your iPad Pro review?!?!?Oldair - Tuesday, January 5, 2016 - link
Since PX2 is 8tflops, and its a dual soc, would the Single tegra SOC be 4tflops, or at least around there, because if so, then that would make it way ahead of today's current gen consoles.psychobriggsy - Tuesday, January 5, 2016 - link
Nope, because there's also two Pascal GPUs (3 TFLOPS or thereabouts each) on board too.extide - Tuesday, January 5, 2016 - link
That 8GLOPS includes the 2 discreet pascal GPU's as well.Yojimbo - Tuesday, January 5, 2016 - link
One thing I think it would be safe to assume is that Denver has been worked on in the 18+ months since Denver TK1, and that the core used in Drive PX 2 is not the same iteration as that used in TK1. It seems useless to speculate whether it has the same performance characteristics of the previous version or not.But even if it did, other things to consider are that Drive PX 2 is a compute-oriented setup, and Pascal supports unified memory. The Drive PX 2 system is designed for the particular workload of machine learning for self-driving cars. The type of stuff going to be run on these machines is well known, so they should be able to much more finely tune what gets executed by what, I am guessing?
Finally, unlike last year with the Drive PX where NVIDIA said Drive PX was powered by two Tegra X1's, NVIDIA didn't announce an SOC name being used with the Drive PX 2. That seems like an indication that this particular SOC configuration isn't planned to be marketed outside of Drive PX 2 or similar in-house systems.
eddman - Tuesday, January 5, 2016 - link
It's probably as you wrote. It is quite possible that their phone/tablet variant of Parker will be different from the one in PX 2.vladx - Wednesday, January 6, 2016 - link
Nvidia dropped out off the phone market entirely.lucam - Wednesday, January 6, 2016 - link
and it also seems it's getting away from tablets too focusing entirely on automotive.psychobriggsy - Tuesday, January 5, 2016 - link
I think we're looking at 384 CUDA cores at around 30% higher clocks, to achieve around 1 TFLOPS FP32.And the Pascal GPUs are 3 TFLOPS (likely verified by 24 DLTOps figure).
3 + 3 + 1 + 1 = 8 TFLOPS - the number quoted for the device overall.
extide - Tuesday, January 5, 2016 - link
Sounds about right to me.phoenix_rizzen - Tuesday, January 5, 2016 - link
From the article:"this new SoC (Tegra P1?) has four Cortex A57s and two Denver CPUs. As of now it isn’t clear whether this is the same iteration of the Denver architecture that we saw in the Tegra K1. However, regardless of what architecture it is we’re still looking at a CPU architecture that is at least partially an ARM in-order core with a wide, out of order VLIW core that relies on dynamic code optimization to translate ARM instructions into the VLIW core ISA."
The Cortex-A57 CPU is an out-of-order CPU architecture, so the phrase "an ARM in-order core with a wide, out of order VLIW core" seems out-of-place and incorrect.
Krysto - Tuesday, January 5, 2016 - link
Nvidia seems to always be behind one way or another. Denver is a 2014 core, and A57 is a first half of 2015 core. This will will arrive in the second half of 2016 most likely, and will have to compete against Kryo and A72.I blame this on Nvidia's lazy execution. They'd rather keep being lazy and use an old core, instead of running a tighter ship over there and launch a more contemporary CPU like everyone else.
Yojimbo - Tuesday, January 5, 2016 - link
Denver is a project name, and NVIDIA doesn't seem to use it the way you are using it here by putting it in the same category as the term "A57" and assigning it a particular date (2014). I think it's safe to assume that the 2016 Denver CPU will not be the same core as the 2014 Denver CPU. Besides, speculation on Parker based on Drive PX 2 is dubious being that NVIDIA didn't mention any SOC name when talking about Drive PX 2. Drive PX 2 has particular demands which are different from the demands of a tablet or smartphone device. Even if the Denver used in Drive PX 2 is the 2014 Denver, there is no good reason to assume that NVIDIA's next tablet/console SOC, assuming there is one, will use the same core or resemble Drive PX 2 in any way.