Are CPUs becoming redundant for games?

[Abuse]

Next generation games will demand much more than just fast rendering of triangles and pixels—they will require the GPU to compute physics, simulate artificial intelligence, and render advanced cinematic effects. These demands are all met by the next generation NVIDIA® CUDA™ architecture in GF100 GPUs.

NVidia want to off-load AI to the GPU now too?!
Unless we're talking about genetic algorithms and other computationally expensive ways of implementing learning systems, what possible reason would there be to off-load a few AI scripts to the GPU?!

[Riven]

Marketing speach and lies.

You can make the (CPU) AI smarter by handing it better processed data (GPU).

Besides that, we haven't exactly reached the top of usable processing power. We're just using new techniques to move the load to somewhere else, and fill the gap with something else.

[pjt33]

NVidia want to off-load AI to the GPU now too?!
Unless we're talking about genetic algorithms and other computationally expensive ways of implementing learning systems, what possible reason would there be to off-load a few AI scripts to the GPU?!

Parallel processing could be really handy for things like path-finding.

Games published by our own members! Check 'em out!

[lhkbob]

I remember reading about some techniques to do dynamic map analysis to determine where troops could walk being improved by offloading it to the GPU.  It's not really AI but it's data needed to perform the AI.

[princec]

If you're going to rely on GPUs to do that sort of thing... er... why not perhaps look at that other core sat there probably doing nothing instead...

Cas

[pjt33]

If you're going to rely on GPUs to do that sort of thing... er... why not perhaps look at that other core sat there probably doing nothing instead...

Aren't GPUs supposed to be better optimised for SIMD than an i386 core?

[Roquen]

Most games are CPU and not GPU bound.  So in that case it's a win to move GPU friendly routines to the graphics card and let the CPUs perform other work.  Also the GPU subsystem is much better at certain types of computations and vice-versa.  Using the best suited is reasonable (when possible).

Specifically for AIs, well, they are sometime more complex than a simple state-machine.  As an example: team tactics - a group of n <whatever> working together.  The group has a set of goals and each member as a set of beliefs, desires and goals.  So stuff like path planning, "what do I perceive", etc queries are part of the picture.  Also there are the various 'learning' methods which require a lot of number crunching.

[Abuse]

Most games are CPU and not GPU bound.

I'd contest that statement.

I have only come across 1 AAA game that has appeared to have been CPU limited on my Q6600 (quad core, 2.4ghz), and that was M$ Flight Sim X.

I can say this with some confidence, because I find that enabling SLI on my 2x8800GTX setup (effectively increasing GPU capacity by ~50%) improves framerates in every game that is supported by the drivers.

This includes all of the past resource hog games like Crysis & GTA IV.

In the case of M$ Flight Sim X, the game is renowned for being badly written - failing to take any advantage of multiple cores or gpu horsepower. Single core @ 4ghz is the only way to get a decent framerate.

[Orangy Tang]

Most games are CPU and not GPU bound. 

I strongly disagree. Most games are fill rate bound these days, and that's only going to get worse as post processing effects get increasingly sophisticated. As a random example, iD moved all of their sound processing for Rage off dedicated sound hardware and onto the CPU because they found that they had CPU time to spare and found that hardware issues weren't worth it.

Most games (even professional ones) are *still* on an old-school single-threaded approach because CPUs are fast enough and it works. We've got spare cores coming out of our ears that are (largely) dormant in all but the most cutting edge of games engines (ie. those that have been written with consoles in mind, where the actual clock speeds are lower).

I'm with Riven on this - marketing BS. They want you to think 'faster AI == smarter AI == better game', but fast, dumb AI is still just dumb AI.

[Hansdampf]

In the not so far future there won't be a difference between gpu and cpu and everything will be raytraced.  At least I hope so, easier to code.

Games published by our own members! Check 'em out!

[Roquen]

It's unimportant, so I don't want to belabor this point (well TOO much)  But: simple tests likes changing the value of your CPUs overclock rate and enable/disabling SLI and seeing a change of performance is a couple of sample points.   However did doesn't really tell you if your either CPU or GPU bound because correlation does not imply causation.  You really need embedded performance tools (tracing drivers, etc) to determine this, along with knowing how the system is constructed.  As a counter example: a poorly constructed design will have alternating idles between the two sub-systems and is therefore bound by poor design rather than by either CPU or GPU.

Most games (even professional ones) are *still* on an old-school single-threaded approach because CPUs are fast enough and it works. We've got spare cores coming out of our ears that are (largely) dormant in all but the most cutting edge of games engines (ie. those that have been written with consoles in mind, where the actual clock speeds are lower).

I'd argue that it's nearly impossible for a single threaded approach to keep the GPU in full utilization.

Why not perform fullscreen effects on the CPU instead of the GPU?  Ignoring even the cost of data motion, it would be silly..the GPU performs these kinds of computations much, much faster.  If your not GPU bound? Why not think about moving GPU friendly routines off the CPU and do somthing else with the resource.  Physics on the GPU was marketing hype...now it's real.

[princec]

I can't quite see that GPUs are going to be any good at AI. Especially as the results of their computations need to come back to the CPU, and the bus design etc. is very specifically crap at that kind of thing.

Cas

[Riven]

I can't quite see that GPUs are going to be any good at AI. Especially as the results of their computations need to come back to the CPU, and the bus design etc. is very specifically crap at that kind of thing.

Cas

Influence maps for the AI. (rendering lots of alpha-blended circles to a texture)

[JL235]

Maybe a little off-topic, but last year there was a range of companies all pushing the idea of interactively streaming games as video from a server where it's running. If it catches on (which I hope it definitely does for hosting demos on developer sites) then it removes the need for all high-end hardware on the client-side.

[Roquen]

Maybe we should merge this thread with the one on closures. 

[h3ckboy]

Maybe a little off-topic, but last year there was a range of companies all pushing the idea of interactively streaming games as video from a server where it's running. If it catches on (which I hope it definitely does for hosting demos on developer sites) then it removes the need for all high-end hardware on the client-side.

yeah my bro told me about that.

[jezek2]

Maybe a little off-topic, but last year there was a range of companies all pushing the idea of interactively streaming games as video from a server where it's running. If it catches on (which I hope it definitely does for hosting demos on developer sites) then it removes the need for all high-end hardware on the client-side.

http://arstechnica.com/gaming/news/2010/01/onlive-demoed-lag-graphics-are-a-problem.ars

[Riven]

And suddenly the developer is going to pay for the hardware too. It's just rubbish technology, from many angles.

[whome]

http://arstechnica.com/gaming/news/2010/01/onlive-demoed-lag-graphics-are-a-problem.ars

Onlive and Gaikai game streaming services are a hot topic atm. Nice x264-dev blog about the use of (almost) lag free video compression. x264-dev developer is (indirectly) saying he provided a working implementation to Gaikai service. He is not grazy about Onlive variation.

It seems to work and should provide an excellent experience on certain game genres, such as RTS, Laser Squad Nemesis and point&click games. Imagine, you at the hotel with a crappy work laptop and want to play few rounds of Starcraft. Gaikai may give you that within a year or two. Only big question is do they find a working business model and subscription fees.

From the publisher point of view, the end of pirating and 2nd hand Gamestop game sells.

[Riven]

Onlive and Gaikai game streaming services are a hot topic atm. Nice x264-dev blog about the use of (almost) lag free video compression. x264-dev developer is (indirectly) saying he provided a working implementation to Gaikai service. He is not grazy about Onlive variation.

It seems to work and should provide an excellent experience on certain game genres, such as RTS, Laser Squad Nemesis and point&click games. Imagine, you at the hotel with a crappy work laptop and want to play few rounds of Starcraft. Gaikai may give you that within a year or two. Only big question is do they find a working business model and subscription fees.

From the publisher point of view, the end of pirating and 2nd hand Gamestop game sells.

What if a games becomes a hit? It simply won't scale well. You'd face massive hardware/network investments to keep your little smash hit game up and running.

[h3ckboy]

it is not he producers who do it, it is a service that you sign up for. a thir party, that is providing the horsepower.

[princec]

And that's the problem - the horsepower. It's going to be really expensive effectively buying not only the rendering hardware - which is essentially a top-spec gaming PC per simultaneous customer, but also the encoding hardware, and the bandwidth to transmit the pictures (although realistically I expect the bandwidth isn't going to be too high as they'll transmit at 29.97fps 720p most likely).

Successful games might have 10,000 simultaneous players - so that's 10,000 top-spec gaming PCs. For one game. Ahahaa.

Cas

[princec]

hmm and also there will be some unavoidable and significant lag so it really will rule out the proper nice FPS games. I knows this coz I work in broadcast. Even playing back local files can incur a 100ms lag, and that's before any rendering, encoding, transmission is done, and before any input lag, and before any client/server lag between the rendering machines and the server machines.

Cas

[Riven]

it is not he producers who do it, it is a service that you sign up for. a thir party, that is providing the horsepower.

Ok, but somebody has to pay it.

Players have perfectly capable hardware in front of them, but *they* have to pay not only to you, but also to the platform provider. In the end, it's simply yet another (high) cost, and every dollar spent by the player on the platform provider, is not in your pocket.

[princec]

It's also worth noticing that this is just another impending trough in the endless thin-client/fat-client cycle. It's happened at least 3 times already in the last 30 years, and I expect this is just another pointless fad before everyone realises that fat-clients are the way to go again. Meh.

Cas

[JL235]

I disagree about cost completely ruling it out. There would be a high startup price but buying, maintaining and running high-end machines in bulk will be far cheaper then doing this indevidually. You can also share the resources between multiple users lowering the cost per user even further.