I will also add that we use the TX1 in our unmanned vehicles all the time and it works great.

EDIT 1, I mean real UAVs like the kinds that fly over 3km high, not grad student projects (no offence, i was once one of them too). We use connect tech for a ruggedized carrier board.

This sounds super cool. I want to know more, do you guys have a web site?

We do not.

>With the dev kit leading the charge for Jetson TX2, NVIDIA will be selling it for $599 retail/$299 education, the same price the Jetson TX1 dev kit launched at back in 2015.

2 years later and the TX 1 dev kit still costs ~500$. Now that pre-orders for the TX 2 are open, shouldn't the price go down significantly? I'd really like to try it but the price is too steep IMO

They are targeting people building autonomous cars and UAVs. Not targeting hobbyists. For someone doing this for real work, this is inexpensive. I am actually surprised it is only $600.

However I'll also say that I'd like one and would be more likely to buy if it was $200. I have a Jetson TK1, but I got that for free when they first came out.

> I have a Jetson TK1, but I got that for free when they first came out.

How did you manage that? Were you attending a conference where they were handing them out or something?

Although I have to say my appetite for niche ARM boards has diminished significantly since I bought the PandaBoard ES and found the support from TI to be very poor. Had to send my first board back for repair under RMA because apparently the HDMI connection cannot be hot plugged (or you damage the board) even though hotplug is part of the friggin HDMI spec.

The TK1 boards (at least) have their own set of problems too. I was at a conference where one of the vendors was showing off a little stack of TK1s he was using as a cluster... but apparently it needed to warm up before the GPU would work properly.

>How did you manage that

They had a program where you submitted a very brief description of a project that you wished to pursue and they would send you a board.

I also have several of these niche ARM boards - collecting dust mostly.

I'm in the same boat.

Apparently education pricing is only $300 though.

I sometimes think that I should take a 0.5 credit online course at my local community college every once in a while so I can abuse those sweet, sweet education discounts.

Last time I was thinking this was actually for the same project. Point Grey makes some sweet-ass USB 3.0 cameras that can do 120 fps, I pump it into a Jetson, distort it for the screen and add overlays, then dump it back out to my Vive. Boom, ghetto augmented-reality setup.

It's a tricky project in a technical sense in several respects - you need wide angles plus decent resolution plus fast framerate and low response times. The really nice thing about PointGrey's cameras is not only are they fast and hi-res, they have plain old C-mounts, so I could use real superwide TV lenses instead of ghetto crap I tore out of webcams. The downside is, they retail for like $700 a pop or something.

I was in the same boat for the PINQ board. I do support students getting things cheap but I really wonder if gouging engineers/scientists/hackers who don't need the hw for a current work project makes any sense.

Does the TX2 boards have gpio connectors? I've been meaning to upgrade my rasberry pi robot into a more serious unmanned vehicle.

Yes according to the spec here:

http://www.nvidia.com/object/embedded-systems-dev-kits-modul...

I've recently tested Unreal Engine demos on the TX1 and must admit the result are really impressive for such a small device.

Based on my comparisons of Maxwell core to Pascal core efficiency, this should results in roughly a 1.4x speed up of FP32.

An NVIDIA representative a week ago told me 2x but that might be a bit high, I'd take it as a cap, and maybe it is the case using tensorRT.

Our NVIDIA rep told us early January that 1.5x was a reasonable boost for us to expect.

If you are memory bound the increased throughput from 64 bit to 128 bit wide LPDDR4 bus would likely get you close to 2x.

yes, you are right if you are memory bound!

Isn't the TX1 almost identical to the nvidia shield tv? So why is it so expensive then? ($599 vs $199)

It's a low-volume (relative to Shield) developer platform

The Shield TV is a loss-leader. It is sold primarily as a gaming platform for Nvidia's subscription based game store.

Has anybody seen any data on how the Denver-2 application cores perform?

No, because approximately nobody could get them until now.

Like the memory bump, not clear how many CUDA cores are there..

256, same as before. But they are Pascal, which, at 14nm, is 2x more efficient than 28nm Maxwell.

It used to be the case that efficiency scaled reliably with feature size like that. Dennard scaling has broken down since the mid 2000s, however.

https://en.wikipedia.org/wiki/Dennard_scaling

Wouldn't it be nice if keras 2.0 came out this week too?