Wasn't this kind of thing implemented by GeForce FX back then in DX9 shader model 2.0? ATi doing the required 24bit as per DX9 and FX can only performn 16bit and 32bit?

So blame NVidia to use old technology for consumers (even if they have new one)

It's not that simple. Like with everything else in GPU technology there are tradeoffs. It's been speculated that the reason for Vega's higher than expected power consumption and die size is related to the redesigned cores for packed math. This was also rumored as to why Nvidia left the FP16v2 cores that does RPM on GP100 off it's gaming cards. It's quite possible that the power/size increase isn't worth the performance increase of having packed math in the first place. Like if I told you Nvidia could add FP16v2 to GP102, but it would use 100w more power and they'd have to drop 250 CUDA cores to fit it, would that be worth it? Yeah you'd get some additional acceleration in games that utilize FP16, but you'd lose performance in every other game and it would screw the efficiency. And I'm not saying that 100w/250cuda is what it would take, I have no idea what it would take - but there exists potential downsides to making the decision to include it. It's not all net benefit. So obviously Nvidia has to weigh that to some degree. I imagine that if it's worth it Volta will get it - they already said Volta has a 50% perf/w increase on it's RPM FP32/16 units. So it's looking like it may get included.

Wasn't this kind of thing implemented by GeForce FX back then in DX9 shader model 2.0? ATi doing the required 24bit as per DX9 and FX can only performn 16bit and 32bit?

Kind of - back then it had two separate cores. Now they can do combined operations on one core. But there are limitations to the type of operations you can do with rapid packed math. Anandtech explained it really well in their GP100 article.

when is the NDA lifted on these cards?:3eyes:

when is the NDA lifted on these cards?:3eyes:

LOL you actually replied in the RV Vega thread on Monday and didn't get the NDA was lifted and the cards paper launched? This thread is just about stuff that RX Vega can do and now drawing attention to coming games that are actually using it and a Futuremark demo for it.

LOL you actually replied in the RV Vega thread on Monday and didn't get the NDA was lifted and the cards paper launched? This thread is just about stuff that RX Vega can do and now drawing attention to coming games that are actually using it and a Futuremark demo for it.

I should be obvious that I mean "reviews"! I cannot find anything related to this topic!

Wasn't this kind of thing implemented by GeForce FX back then in DX9 shader model 2.0? ATi doing the required 24bit as per DX9 and FX can only performn 16bit and 32bit?

Basically yes and remember how bad the image quality was compared to the Radeon 9700 and 9800

You're talking about colour depth not render targets there

So AMD need to sacrifice image quality now to keep up with the competition... Then they're selling it as a "feature". Interesting

I'm with you here. They might see a small bump in fps, but at the cost of image quality in things they think nobody will notice. People will notice, they always do. So many people on the internet claim to see crazy detail in high fps, screen tearing, etc. People will complain and it will do more harm than good to have this "feature".

I'm with you here. They might see a small bump in fps, but at the cost of image quality in things they think nobody will notice. People will notice, they always do. So many people on the internet claim to see crazy detail in high fps, screen tearing, etc. People will complain and it will do more harm than good to have this "feature".

AMD already does this with tessellation and most people don't seem to care. I don't really think it's that big of a deal for optimization on noise generation and whatnot, which is what they plan on using it for. Like yeah there might be a visible difference, but I bet in a blind A/B test you couldn't tell which is the higher precision - they'll just look different. I would imagine that they'd have an option to disable it as well.

From what I have seen in the last few years, if Nvidia is not going to jump on the band wagon of fp16 and support it on a hardware level, then you will only see a small hand full of games with this feature and then it will disappear in the dust. Same like all the other unique features that AMD has created in the last couple of years. Features like AMD true audio, mantle and the list go on... The reality is no support from Nvidia means no future support in games...

Let's sweep all those abandoned Nvidia features under the rug, they don't exist 🤓. TrueAudio Next replaced TA nearly a year ago and it's available on GPUOpen. It and Nvidia's VRW audio will probably get a few games with support and fall off a cliff shortly after. Mantle didn't have much reason to exist with DX12 rolling around. Mantle was also groundwork for Vulkan according to articles. It's on AMD or Nvidia when features are actually dumped from GPUs if a better option isn't superseding what's being dumped. It's on developers when something is freely available, ie. TressFX or TAN, and they aren't using it.

I think a lot of those libraries sound good in theory but then when you go to integrate them in games it's a ton of work on the developers part, for relatively little benefit - especially the ones that are VR centric. Little off topic but I'd honestly like to see a big push in sound for video games again - a lot of newer games shipping don't even have proper 5.1 support, it's kind of annoying. Music Producer Benn Jordan (The Flashbulb) recently talked about working with a AAA dev (I don't know if he announced the company publicly) about a project he's working on where all the sound in the game is generated dynamically from a built in modular synthesizer tied to the actions in the game. The way he described it was really cool. UE4 recently added an entirely new sound system as well which also features a modular synth, here is a good video on it: https://www.youtube.com/watch?v=ErejaBCicds Would be cool to get some of this stuff accelerated on GPU's if possible.

I was expecting x87 for Elder Scrolls 6, but now we might even get FP16! The excitement!

So AMD need to sacrifice image quality now to keep up with the competition... Then they're selling it as a "feature". Interesting

I take it you've never done software development before, particularly with C? This isn't going to affect image quality - this is so the GPU doesn't have to work as hard for more basic calculations. It's a similar concept to defining variables as "uint_8" vs "long long". If your number is only going to be a couple digits, you could use "long long" but you're being really wasteful with your resources. With most of us being gamers, a lot of us aren't aware how big of a performance difference something like this can make. For example, of the gaming GPUs that offer FP64 (double precision floats), most of them are around 1/16 the performance of FP32, where some good models will be 1/8. GPUs like the Titans, Vega FE, Quadros, and high-end FirePros tend to offer 1/4 or better. What I'm getting at is increasing precision (whether you actually utilize it or not) hurts performance, and you have to pay extra to lessen the blow. That being said, if much of the workload becomes "half-precision", we shouldn't notice any quality decrease, but a hefty performance increase. I'm guessing FP16 will be useful for things like particle physics and rendering a lot of geometry far in the distance. The only thing I find especially stupid is how AMD is making this out to be some novel idea. This should've been done YEARS ago, and companies like Google are actually what popularized it. A lot of hardware for things like AI use FP16, because mimicking human intelligence works best when fast and approximate, whereas FP32 is (in comparison) too accurate and slow.

I think a lot of those libraries sound good in theory but then when you go to integrate them in games it's a ton of work on the developers part, for relatively little benefit - especially the ones that are VR centric. Little off topic but I'd honestly like to see a big push in sound for video games again - a lot of newer games shipping don't even have proper 5.1 support, it's kind of annoying. Music Producer Benn Jordan (The Flashbulb) recently talked about working with a AAA dev (I don't know if he announced the company publicly) about a project he's working on where all the sound in the game is generated dynamically from a built in modular synthesizer tied to the actions in the game. The way he described it was really cool. UE4 recently added an entirely new sound system as well which also features a modular synth, here is a good video on it: https://www.youtube.com/watch?v=ErejaBCicds Would be cool to get some of this stuff accelerated on GPU's if possible.

Benn Jordan always amazes me with his talent and creativity. A truly original artist.

I think a lot of those libraries sound good in theory but then when you go to integrate them in games it's a ton of work on the developers part, for relatively little benefit - especially the ones that are VR centric. Little off topic but I'd honestly like to see a big push in sound for video games again - a lot of newer games shipping don't even have proper 5.1 support, it's kind of annoying. Music Producer Benn Jordan (The Flashbulb) recently talked about working with a AAA dev (I don't know if he announced the company publicly) about a project he's working on where all the sound in the game is generated dynamically from a built in modular synthesizer tied to the actions in the game. The way he described it was really cool. UE4 recently added an entirely new sound system as well which also features a modular synth, here is a good video on it: https://www.youtube.com/watch?v=ErejaBCicds Would be cool to get some of this stuff accelerated on GPU's if possible.

If it's heavily geared toward VR I could see it being a problem to implement. Lower latency and more realism is a big improvement to me for anything though, not just VR. It seems cool for sure but I worry it'll suffer the same fate as everything else unless it's easy to implement. Watching now. Agreed on sound in games, it's pretty irritating how it's been going lately. On GPU acceleration, if it happens, I wonder if it would be better off left unadvertised as to avoid getting lampooned before it can go anywhere.

So blame NVidia to use old technology for consumers (even if they have new one)

I don't think it's that simple. Nvidia's architecture is far more specialised than AMD's. I'm willing to bet this is the main reason for Nvidia's reign in the past years. AMD's architecture has grown too complex, some kind of jack of all trades master of none. Segregating some of the hardware between pro and gaming chips allow for more specialisation: pro chips get more hardware meant for higher precision maths, while gaming chips get more hardware meant for geometry and rendering techniques. You do get 'less of everything' and 'more of some certain things' as a consumer though. Whether this is good or bad, I don't know. I don't really care about my card's compute performance to be honest, I simply don't have frequent use cases for it.

AMD already does this with tessellation and most people don't seem to care. I don't really think it's that big of a deal for optimization on noise generation and whatnot, which is what they plan on using it for. Like yeah there might be a visible difference, but I bet in a blind A/B test you couldn't tell which is the higher precision - they'll just look different. I would imagine that they'd have an option to disable it as well.

Nah, people just like to conveniently ignore it. There are differences above 16x factor in a good number of games, yet people are still self pitying about how it doesn't matter: http://imgur.com/a/VorPz These said people tend to be AMD fanatics, as you can probably gather.

If anyone thinks there will be no visible differences, you can find old examples like this onehttp://www.hwupgrade.it/articoli/skvideo/1013/radeon-x800-e-il-momento-di-r420_15.html. I hope all vega owners will have a chance to enjoy 2005. once again...

As someone who likes the idea of FP16, I'm well aware there CAN be visual differences, if this isn't implemented correctly. I feel that having an entire game limited to FP16 is a real crappy way to improve performance. However, if specific calculations within the game can be done in FP16, it could yield perfectly normal results while improving performance. I don't really see why people are looking at this so black and white. As someone who doesn't develop 3D games, maybe there's something I'm missing, but I get the impression this is meant to be a bit more dynamic and not applied to the entire application. I'm guessing AMD allows custom application profiles where you can force-enable it if you want, which they may have done for the games mentioned in the article.

Pretty sure the PS4pro can do this. http://gamingbolt.com/mass-effect-andromeda-ps4-pro-dev-explains-benefits-of-checkerboard-rendering-30-improvement-due-to-fp16

it will disappear in the dust. Same like all the other unique features that AMD has created in the last couple of years. Features like AMD mantle

mantle turned into vulkan which is very much alive.

So AMD need to sacrifice image quality now to keep up with the competition... Then they're selling it as a "feature".

as you later mentioned, nvidia coded anisotropic filtering optimizations into their drivers with no option to disable it, resulting in blurry subpar textures at any angle even @x16AF. but seriously :P both companies learned from that debacle & now have driver-based optimizations that are toggleable. well, this is taking optimization one step further (into hardware) & id be shocked if you can tell the difference even from a static screenshot of the exact same scene of fp16-optimized vs fp32 noise filter effect, for example.

It's been speculated that the reason for Vega's higher than expected power consumption and die size is related to the redesigned cores for packed math. This was also rumored as to why Nvidia left the FP16v2 cores that does RPM on GP100 off it's gaming cards.

considering gp100 is so """"small"""" i would imagine this certainly was taken into consideration, perhaps testing the waters with the plan to integrate a similar feature on-die for volta

AMD already does this with tessellation and most people don't seem to care. I would imagine that they'd have an option to disable it as well.

exactly.

However, if specific calculations within the game can be done in FP16, it could yield perfectly normal results while improving performance. I don't really see why people are looking at this so black and white.

cos people who dont think are usually the first to speak...& sometimes the loudest 😀