Jetson TX 2

[bentomo]

Not sure if anyone else has posted this yet. Tegra X2 coming development coming soon. Too bad it's still pricey.

http://www.nvidia.com/object/embedded-sy...dules.html

They say making a breakout board will be much easier for this one though.

Just from looking at the webpage specs it doesn't look like there will be much of a difference between TX1 and TX2 other than the Pascal change that could be useful for dolphin.

[degasus]

The big change for dolphin are those two Denver2 CPU cores in it. The pascal GPU won't matter.

[bentomo]

Ah my ignorance is showing. A lot of dolphin will have to be written specifically for the denver chiplet won't it?

[degasus]

No, it's also ARMv8, but the Denver has a different performance than the A57, and we're usually bottlenecked by the CPU.

[Nintonito]

The Denver and A57's run in a very weird parallel, with the A57's handling instances where the Denver cores are repeatly forced to convert instructions. Using A57's has the unfortunate problem of not improving performance over the X1, while the Denver 2 cores are potentially faster than the Denver 1. Another big issue is that when both clusters are active, all cores are capped at 1.4GHz for power reasons, which basically eliminates the possibility of better performance in those scenarios. The design of the Parker chip is altogether extremely odd, with Nvidia basically opting to use a quad A57 cluster instead of an ARM ucode translator like on the K1.

[degasus]

Oh, the Denver recompiler is written in ARM itself and run on the A57? This means we should be able to get our code on the Denver ourself

[Nintonito]

Oh, the Denver recompiler is written in ARM itself and run on the A57? This means we should be able to get our code on the Denver ourself

That's the general impression. I cannot confirm this as it's really under wraps, but early testing in the Drive PX2 indicates something along those lines. Using actual arm scores instead of a translator to fill the weakness of the first gen design.