Would you buy a TPU?

[Gregorovitch]

One interesting side issue with this Alpha Zero business and it's relevance, if any, to mainstream gaming is the idea that you can't run deep NNs on home PCs as part of a markertable quality game, or if you tried it would run at 1FPM becasue the NN would simply hog every ounce of power your CPU/GPU had.

So Alpha Zero runs on Tensor Processing Units. These things are direct analogues of Graphics Processing Units except instead of crunching a gazillion floating point sums in parallel they crunch gazillions of 8-bit integer sums in parallel which is all NNs need for calculating weights and is orders of magnitude less processor intensive than floating point calulations. In fact pre-TPUs GPUs where a popular choice for running NNs (for the same reason they are for mining bitcoins).

You can see a picture of a TPU in this article:

http://www.electronicdesign.com/industrial/cpus-gpus-and-now-ai-chips

We can see the familiar PCI slot interface, but the unit is obviously a bit bigger and chunkier than a GTX1070, looks like you'd need some sort of mamoth case to house it. However one imagines this is an industrial strength one designed for Google's data centres and that a domestic version suitable for running a game AI would come in significantly smaller.

The implication of this would seem to be that with a TPU intalled alongside a GPU in your PC a game using it for a NN based AI would actually run faster since the CPU would no longer have to run a significant proportion of it's AI calculations on the CPU.

Now suppose that Codeforce announced that Distant Worlds 2 would be shipping with an optional NN AI usable by players who had a TPU installed (and lets assume by 2019 nVidia, AMD, Intel and possibly Google themselves have made available consumer grade TPU versions):

Would you buy one?
How much would you pay for it?

[Nasarog]

Not I. I am not a trail blazer.

[BlackSmokeDMax]

I would only buy one depending on the price. If it's a cheap gamble... maybe or maybe even probably. Otherwise, I'll be waiting to see how the market for it shakes out.

[marlowe221]

I have no idea what any of that means...

[AK_icebear]

No. I doubt anyone would.

[Mr.Kill]

"8-bit integer sums"

That's freaking worthless for my gaming needs XD

Count me out.

[athelasloraiel]

I would

[Greygamer]

Originally posted by marlowe221:

I have no idea what any of that means...

Me neither, as a side note I'm not even interested in VR

[Mansen]

It's definitely an interest concept - But I've never been an early adopter, since it always tends to be a bad investment.

[Hans Lemurson]

It would certainly be interesting, to have a dedicated "AI Chip" on a computer, but I have a hard time imagining how they would train up a Neural Network to play these highly complex games.

Neural Networks take a LOT of cpu-time to create in the first place, and are evolved from countless iterations of gameplay. Resolving a single turn of a game of Chess or Go is an utterly trivial task, computation-wise. You can breeze through those games with almost no effort. Thinking the moves through is complicated, sure, but the games are simple to compute.

Strategy games involve crunching multiple spreadsheet's worth of data each turn. If you disregard the costs of AI calculations and just look at the game's internal number crunching, then even something like Civ6 will run pretty fast, but it's still orders of magnitude slower than Chess or Go. Anything less than 1/10th of a second feels "instant" to most people, but there is a huge gulf between something that takes 100 milliseconds to compute and something that take 1 microsecond.

I think it's going to be a long long time before there's anything useful that a home user would want to run on a Tensor Processing Unit.

[Emperator Orionii]

Originally posted by Gregorovitch:

One interesting side issue with this Alpha Zero business and it's relevance, if any, to mainstream gaming is the idea that you can't run deep NNs on home PCs as part of a markertable quality game, or if you tried it would run at 1FPM becasue the NN would simply hog every ounce of power your CPU/GPU had.

Citation needed.

How many neural connections does Alpha Zero has, how long would it run on some (specs please) home PC? Unity3D engine hogs CPU, GPU, RAM and makes absurd amount of HDD traffic and yet people are using it home PCs regularly.

[Sithuk]

Wouldn't a cloud based AI be more likely in the near term? Why doesn't Google monetise a neural net (NN) AI with a subscription model? The easiest way is likely to be support for games with multiplayer support first, as the NN AI can be the non-player character(s).

So GalCiv3 would more likely have a NN AI option before Distant Worlds, which is single player only.

[Icemania]

Originally posted by Gregorovitch:

Would you buy one?

I would rather contribution towards a sponsorhip for a sharp AI PHD student to do a project focused on adapting AlphaZero to a good 4X game and provide it's even feasible first.

[Emperator Orionii]

Originally posted by Icemania:

Originally posted by Gregorovitch:

Would you buy one?

I would rather contribution towards a sponsorhip for a sharp AI PHD student to do a project focused on adapting AlphaZero to a good 4X game and provide it's even feasible first.

I'd work on that PHD :)

[Nasarog]

Originally posted by Emperator Orionii:

Originally posted by Icemania:

I would rather contribution towards a sponsorhip for a sharp AI PHD student to do a project focused on adapting AlphaZero to a good 4X game and provide it's even feasible first.

I'd work on that PHD :)

Who is funding this PhD? eXplorminate? We'd need some more Patreon support before we commit.

[Gregorovitch]

Erik and Elliot :)

BTW, WTF is going down with this tick box quote thing - what's it supposed to do?

"You are about to mark this post as the answer to this thread. This will indicate that the original post has been answered and link to this specific post. You can remove this or indicate a different post as the answer at any time if you change your mind."

Er, what?