Nividia TX2 (ARM)


The Nvidia TX2 is a $100 more than the TX1 and it include 4-core A57 running at 1.6MHz and 2-core Something Nvidia, full SD card, and 8 Gb of memory in the chip and runs at 15 amps; 7.5 amps if you are doing anything.
there is already a thread about the Pi3 Reviving the HiFiPi

Now in Opensimulator the rule was 1 region/1 core and all the memory you could get. Now HiFi uses 32km which is about 1900 miles, which is a lot. UE4 Open Worlds is at 100 square miles so I am missing something.

UE4 and Blender doesn’t not run on ARM last time I checked. Blender is okay running on a Intel but with UE4 I don’t know how to do it for ARM and have a full blown HiFi or multi HiFi. What I am looking for animated worlds that you can stream and has qualities like in UE4 but low power. I don’t expect user interaction except maybe some camera stuff.

So what do I do: Nvidia TX2 and have it blow up all the time, or Pi3 and have it blown up some of the time?


doing nothing instead of doing anything. ack.


the TX1 is the same thing but only has 4 GB or memory and doesn’t have the 2-core Nvidia something. TK1 is about $300 and it only has 2 GB of memory but it has a full SD instead of the micro-SD and run on 32-bit. The Nvidia Shield has TX1 but only 3 GB and you got to get rid of the Android but the same price as TK1. Of course Pi3 is so cheap you can build multiple of these things but 1 GB of memory.

TX is 64-bit


well I have been gone for while. first there is no Linux version of HiFi. you have use something like this:
Ubuntu 16.04 (Xenial) Package

Well the good news is the TX2 has Ubu 16.04 already and it have 8 GB. But it only has 256 cores-- Pascal based. From a programming point view, I don’t think Pascal or Maxwell graphics means much. I don’t know what Pi3 has for graphics but with 1 GB is hard just to fit the server on there.

There is a price different. TX2 is $600, Pi3 is $40. There should be a minimum to run HiFi not the typical case. Even UE4 doesn’t require you to have a 970; I was getting by with a 770.

So the 32km is a footpad if you have enough memory. I have seem Opensimulator with lots and lots of objects but you have to have 32 GB. I think its similar to HiFi-- you have to have a lot of memory for 32km.

This is expensive software.


Apologies your thread structure is a bit hard to read but ill try to give some information.

You should be able to run High Fidelity with most graphics card and os that supports the OpenGL 4.1+. the RSP3 simply has a Broadcom VideoCore IV as its GPU, but generally only the CPU is used When hosting a server. Same would apply for the TX2.

Now generally speaking to render in minimum requirement High Fidelity is based on VR minimum requirements, currently is, which generally is accepted to be anything that can run rendering of medium-intensive software at 90 fps. However, you can use slower GPU if you do not need to reach that 90 fps target. Infact, it still runs on a Intel Iris HD and technically a HD4400.

When looking at this way, you could get the High Fidelity Interface client running on a TX2, but not at full 90fps.

They are also working on an Android version which, natively runs on an (ARMv7-8). Currently however you would have to build your own, as there is no ready made build available, and most RSPI experiments are just Server Hosting, not rendering.

Now unlike OpenSim or Second Life, none of the physics are done on the server side as of the moment. So
Most of the entities in question will be static from the perspective of the server, where as dynamic objects are entities updated by the clients connecting.

This is much cheaper to keep track of than have multiple dynamic objects and should be much cheaper on the ram from the server’s perspective.

The 32x32x32 km also is a theoretical limit. You technically could go beyond that but you will start having more noticable floating point issues when trying to align things. It would make more sense to split up areas of a domain if you have such a large area and have a very high entity density in that 32km cube. This is however what they had in the road map a few years back, but connecting multiple domains together is something that hasnt been discussed much since.

You are also most likely keep track of much less objects, as there is no scale limit to the size of models. Most of the ram will be used to Providing assets, doing short caches of voice comms among other.