NVIDIA Tegra 'Parker'

[Sweepr]

NVIDIA Tegra Parker Unveiled

- 2x Denver2 cores + 4x Cortex-A57
- Significant perf/watt improvements
- Highest performance ARM CPU
- 256 Pascal-based CUDA cores
- 50 GB/s bandwidth (twice Tegra X1's 25 GB/s)
- TSCM 16FF

www.computerbase.de/2016-08/nvidia-tegra-parker-denver-2-arm-pascal-16-nm

 

[ShintaiDK]

The Nintendo SoC?

 

[Phynaz]

The day I believe an Nvidia CPU performance comparison chart....

Isn't here yet.

 

[Arachnotronic]

The day I believe an Nvidia CPU performance comparison chart....

Isn't here yet.

The most impressive CPU in NVIDIA's CPU performance chart is the Apple A9X/Twister by far. It is a dual core matching 8 core A72+A53 parts. Qualcomm's Kryo gets utterly wrecked.

 

[poofyhairguy]

Does Pascal have more power per Cuda core than Maxwell?

 

[NTMBK]

The Nintendo SoC?

You really think that thing is going to fit into a handheld, with 6 "big" cores and a 128-bit memory bus? 0_o I would expect a semicustom chip which stripped out a ton of the automotive stuff and had a small CPU cluster.

Does Pascal have more power per Cuda core than Maxwell?

It generally has a higher clock, so yes.

 

[Arachnotronic]

It generally has a higher clock, so yes.

Better memory compression, too. Should be very helpful when we're talking about an SoC with only 50GB/sec bandwidth for six CPUs, GPU, and a bunch of other clients sitting there with non-trivial bandwidth needs.

 

[ShintaiDK]

Does Pascal have more power per Cuda core than Maxwell?

Yes.

 

[NTMBK]

Better memory compression, too. Should be very helpful when we're talking about an SoC with only 50GB/sec bandwidth for six CPUs, GPU, and a bunch of other clients sitting there with non-trivial bandwidth needs.

Ah, I was just thinking about the actual shader cores, not the memory subsystem. Yes, the improved compression would be a great help- though I wouldn't be so sure that Tegra X1 didn't have a similar system in place already. It's a 20nm "Maxwell" GPU which is slightly different from desktop 28nm "Maxwell", having stuff like FP16 SIMD instructions.

I wonder how long ago this design was finished, given that it still uses A57 cores and Parker was originally meant to launch in 2015.

 

[swilli89]

You really think that thing is going to fit into a handheld, with 6 "big" cores and a 128-bit memory bus? 0_o I would expect a semicustom chip which stripped out a ton of the automotive stuff and had a small CPU cluster.



It generally has a higher clock, so yes.

I do, clocks/power consumption can be tweaked to the proper level and the rumored Nintendo part has to do double duty as a home console connected to a TV.

 

[tipoo]

If the full 4 A57 + 2 Denver 2 configuration makes it to the NX, I'd definitely give the CPU side of that a slight nod over 6.something Jaguar cores.

 

[tipoo]

You really think that thing is going to fit into a handheld, with 6 "big" cores and a 128-bit memory bus? 0_o I would expect a semicustom chip which stripped out a ton of the automotive stuff and had a small CPU cluster.



It generally has a higher clock, so yes.

It's a tablet form factor, and phones already absorb similar core layouts.
The PS Vita had 128 bit memory years ago in a smaller form factor than the NX.

What makes you doubt that's possible? Tiny phones have 4 A72s. And it's also likely it'll clock lower when mobile than when docked.

 

[imported_bman]

One of the slides says it supports e-MMC 5.2, unannounced update to the e-MMC standard?

 

[jpiniero]

Only 256 cores is pretty disappointing if the NX is using. No way would this be good enough to be even close to the XBone. Probably somewhere in the 600-650 GFLOPS range. The Drive is probably using the GP106-300 as it's dGPUs, with the clocks dropped a little bit compared to the 1060 3 GB.

 

[poofyhairguy]

Only 256 cores is pretty disappointing if the NX is using. No way would this be good enough to be even close to the XBone. Probably somewhere in the 600-650 GFLOPS range.

That is still double a Wii U right?

 

[jpiniero]

That is still double a Wii U right?

More or less. The Wii U's GPU is similar to the 5550 (320 cores at 550 Mhz)

 

[dogen1]

That is still double a Wii U right?

Wii U is only ~176GFlops, and used AMDs older VLIW arch(iirc, max utilization is only around 80%?) Hypothetically ,if nx is ~600 it could be above 4x in real world performance.

 

[poofyhairguy]

Wii U is only ~176GFlops, and used AMDs older VLIW arch(iirc, max utilization is only around 80%?) Hypothetically ,if nx is ~600 it would most likely be above 4x in real world performance.

Wow add in a CPU upgrade and you can see why Nintendo was happy with Tegra.

 

[Exophase]

The most impressive CPU in NVIDIA's CPU performance chart is the Apple A9X/Twister by far. It is a dual core matching 8 core A72+A53 parts. Qualcomm's Kryo gets utterly wrecked.

But to be fair, it's a tablet (with tablet-oriented SoC) being compared with phones that have lower thermal headroom. A9 might be a better comparison.

Looks like there's something about Qualcomm cores doing especially badly on SpecInt2k. nVidia loves beating them up over this, I remember when they gloated about Cortex-A9 having higher IPC than Krait in this.

When you get down to it, it's really telling that this is the one and only bench nVidia shows.. a very (unrealistically) parallel benchmark that they can win with brute force having enough "big" cores, most likely with no thermal constraints whatsoever. Contrast with the publication of the first Denver in K1, where they listed several benchmarks.

 

[dark zero]

The most impressive CPU in NVIDIA's CPU performance chart is the Apple A9X/Twister by far. It is a dual core matching 8 core A72+A53 parts. Qualcomm's Kryo gets utterly wrecked.

The power consumption will be totally insane...

 

[Headfoot]

I wouldn't doubt that it would totally power off one of the clusters (probably the a57's?) and heavily downclock in mobile mode, relying on the wide low-clock Denver cores. You can aggregate many threads to fewer cores, just not the other way around. So developers would target a full 6 core workload while writing the code and it would get aggregated in mobile mode onto those Denver cores which are especially wide and especially good at getting ILP out.

Since Denver needs software to optimize the workload and extract maximum parallelism from a workload, I would imagine that they're implementing a cache of some sort for these compiled/extracted workloads to further reduce power

 

[2blzd]

Hello Nintendo NX!

 

[FIVR]

Nintendo would be much better off just building some kind of proprietary dock for an iPad Pro and coding for iOS. They could then sell their games (as cheaper versions, "mobile" alternatives) in the app store to anybody with an iOS device and enable the full games to be played when the dock is attached, perhaps adding to the experience with their trademark Nintendo gimmicks and decent ergonomic and kinesthetic design prowess, exemplified by the N64 and Wii controller designs.

The revenue from iOS would be massive, and the iPad Pro already has a far superior cpu/gpu to any Tegra, including Parker. The iPad Pro already has 50+ Gbps memory bandwidth, it also has a manageable TDP and has been tested not to throttle (as all Tegra chips are notoriously prone to high power budgets and throttling).

This is nothing short of an idiotic move by Nintendo, but they are a company that has long been notorious for ill-conceived hardware decisions. Case in point, shoving some kind of 12 watt TDP automotive part they probably pulled out of a Tesla scrap yard into a "handheld" or tablet and expecting it to perform. They will down clock the %*#@ out of it and it will still draw too much power and perform like garbage, expect about what you got in the WiiU.

 

[dark zero]

RIP Nintendo on consoles... No one will miss you... Specially with those A57 cores who are really hot chip and needs Active Cooling

 

[Sweepr]

Eurogamer and DigiTimes mentioned Tegra X1 for the Nintendo NX, now another website says it's based on Tegra X2:

Not too long ago, we heard from Eurogamer that Nintendo NX is using the Nvidia TEGRA X1 mobile chip. It is a relatively old chip and didn’t pack enough punch to be considered anywhere near as powerful as current-gen machines.

However, sources close to the project claim that Nintendo NX is indeed using a TEGRA Chip but not the X1. The company is partnering with Nvidia to use its upcoming TEGRA X2 that is more powerful compared to the original model.

http://segmentnext.com/2016/08/17/report-nintendo-nx-is-using-tegra-x2-chip-more-powerful-than-x1

This will be fun to watch.