I didn't bring it up I came into this thread & my name was in it .
-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
Wolfenstein II and Far Cry 5 will support FP16 Rapid Packed Math
- Thread starter ethomaz
- Start date
I didn't bring it up I came into this thread & my name was in it .
Zombie James
Banned
Anything that helps boost performance and efficiency is welcome. Checkerboard rendering, this FP16 packed math stuff, whatever. It's all good, imho.
blastprocessor
The Amiga Brotherhood
This is good news for PS4 pro gamers. Squeeze more performance out of the box 📦
There have always been games that implemented new technology, often with the help of manufacturers to try to demonstrate the new tech. Happened with 3dnow many years ago.
This doesn't mean anything, though. It would be something if it would be implemented in all major engines, though. But like with asynchronous compute, even mentioned in this article, we will see
a) if it will be implemented
b) when it will be implemented
c) what the exact benefits will be
This doesn't mean anything, though. It would be something if it would be implemented in all major engines, though. But like with asynchronous compute, even mentioned in this article, we will see
a) if it will be implemented
b) when it will be implemented
c) what the exact benefits will be
stan423321
Member
I have no idea how does ambient occlusion calculation work, but based on that no idea, 1: it would work calculation wise, 2: it would quickly get limited by memory bandwidth anyway.
Are lower-precision results from FP16 useful for "physically" distant elements? With real-world optics, every time a distance is doubled, you're quartering the visual info (assuming the inverse-square principle). Can these calls be used for low LOD elements, or is that basically hogwash?
D
Deleted member 325805
Unconfirmed Member
Yup, nothing kills a game quicker for me than performance issues, be it low frame rate, inconsistent frame rate, frame pacing errors, asset load hitching or pop in.
InsaneTiger
Member
Agreed
Polygonal_Sprite
Member
The way people are attacking you without you even being in threads about this subject is really out of order imo. Isn't one of the core rules of this forum to attack the post not the poster when disagreeing?
People are acting like you said PS4 Pro games will outperform Xbox One X games just because you stated the fact that on paper, the PS4 Pro GPU is 8.4TF @ FP16.
Several developers from major companies have already said on forums that FP16 now makes up as much as 30% of their code in certain modern games.
Again look at Snake Pass, a game that uses FP16 and has no business running on Switch in it's handheld mode. This modern UE4 game only runs on Switch because of FP16.
If all FP16 does is keep PS4 Pro / Xbox One X games above 30fps then great!
OnQ gets piled on because he consistently and repeatedly spreads FUD.
Like your 30% claim. I've never seen any quote from a developer actually making the claim that 30% of all their games code is fp16. I believe I have seen a developer use those numbers when specifically mentioning their checkerboard method.
The PS4 pro will perform exactly like you'd expect a 4.2 tf machine with a jaguar CPU would perform.
Polygonal_Sprite
Member
I never said that. I said "in certain modern games". Again people changing what people have said to suit their argument.
Show me the quotes from "several developers from major companies".
As far as I know, one developer/team has said that fp16 has made their checkerboard rendering method 30% more efficient. No developers have claimed that fp16 has turned the pro into a 5.5tf machine
Receipts?
Dictator93
Member
It references the DICE / frostbite presentation, where fp16 RPM sped up 1 part of zhe checkerboarding pass by 30%. Not he whole renderer, just one part of one process. 'Tis fud what he doth speak.
I'm not too surprised to see idTech making use of FP16, as Bethesda seem to have a very close relationship with AMD these days, but FC5 is quite interesting. Of course the Far Cry series uses its own engine, rather than Anvil or Snowdrop, so I wouldn't necessarily assume any wider support from Ubisoft, but it's a start. If the big publisher engines like Anvil and Frostbite start using it it could change the value proposition for Vega by quite a bit, but I'd hold my breath until we actually get any announcements (and benchmarks, for that matter).
As far as I can recall the 30% figure came from seppi over on beyond3d, but I think it was more a general estimate than relating to a specific game. He used to work on the Trials games and his current project is a semi-raytraced claymation game which looks amazing but isn't really representative of a typical game engine.
Regarding game engine support for FP16, the mobile branch of UE4 uses FP16 for all pixel shaders, although what proportion of shader workload is pixel vs geometry vs compute I don't know (although it would likely vary quite a lot from one game to the next).
Judging by the performance differential between Switch/PS4/PS4Pro in Snake Pass (the only UE4 comparison we have across the three platforms) it would appear that they're making quite a bit of use of FP16 in compatible consoles as well (or at least allowing developers to use it), but obviously it's impossible to say how extensive it is without a direct quote from the devs.
As far as I can recall the 30% figure came from seppi over on beyond3d, but I think it was more a general estimate than relating to a specific game. He used to work on the Trials games and his current project is a semi-raytraced claymation game which looks amazing but isn't really representative of a typical game engine.
Regarding game engine support for FP16, the mobile branch of UE4 uses FP16 for all pixel shaders, although what proportion of shader workload is pixel vs geometry vs compute I don't know (although it would likely vary quite a lot from one game to the next).
Judging by the performance differential between Switch/PS4/PS4Pro in Snake Pass (the only UE4 comparison we have across the three platforms) it would appear that they're making quite a bit of use of FP16 in compatible consoles as well (or at least allowing developers to use it), but obviously it's impossible to say how extensive it is without a direct quote from the devs.
FunkyDealer
Banned
I'd like to know why that is. Isn't it a weird feature to drop?
B-B-Bomba!
Member
See I just knew that math packing was too damn slow these days
I'm not hardware proficient, at least not as much as a developer. But it looks like both platforms have taken somewhat different approaches in how they want their exclusives games rendered? Am I translating that right?
Metalingus5150
Member
Doesn't really seem like they need it to get that extra performance out of the xbonex gpu.
Cerny knows his shit. FP16 + Sony tech savvy first party dev working exclusively will produced some magical results.
No such thing, more performance is more performance.
They didn't drop it they never had it
Trogdor1123
Member
So how many rams does it download for me? I assume like 3 at least.
ethomaz
Banned
They didn't drop off anything because Polaris doesn't have double rate FP16.
What probably happened is that Sony asked (and paid) to AMD implement this Vega feature on Pro custom Polaris because they found potential in the feature. MS probably didn't see this potential and chose not implement (paid) it for Scorpio custom Polaris.
Exactly!
Sony basically made a 4K version of the PS4 out of what would have been a slim PS4 but instead of making a slim they used the die shrink to fit double the GPU into a PS4 that stayed about the same size & clocked it a little bit higher. MS basically is making a ready made UWP multimedia computer that's over 4X the power of Xbox One for a straight forward jump for pushing Xbox One games to 4K or close to it while PS4 Pro was made for pushing PS4 games to 4K or close to with cheaper rendering techniques. both will use cheaper rendering techniques when needed but Xbox One X is more of a brute force approach to having 4X the resolution of the base console.
Negotiator
Member
FP16 was used for pixel shaders, but vertex shaders were always FP32 (vertices really need the increased accuracy to prevent artifacts). Back then, we had separate pixel/vertex shader pipelines, so different precision made sense. ATi (AMD) had FP24 for pixel shaders, which was a good compromise between FP16 and FP32.
When the industry migrated from DX9 (discrete shader pipelines) to DX10 (unified shaders), engineers had to adopt FP32 universally because of the vertex shader requirements for high accuracy.
Nowadays we have a new type of shaders: Compute Shaders. These type of shaders don't necessarily require 32-bit precision (i.e. AI pathfinding). nVidia even supports INT8 (8-bit integer) precision in HPC for deep neural network applications. People who compare this with the GeForce FX fiasco are really ignorant, since Compute Shaders didn't exist back then. GPUs were merely toys for 3D graphic processing/video games. Not anymore though:
https://www.theregister.co.uk/2016/09/13/nvidia_p4_p40_gpu_ai/
https://petewarden.com/2015/05/23/why-are-eight-bits-enough-for-deep-neural-networks/
Mixed FP16/FP32 coding will become the norm in the future, since next-gen consoles (PS5/XB2) are going to support it as a baseline feature (most likely with a Navi GPU). RPM, CBR and Ryzen will allow us to have more 4k60 games in the future. Good times ahead.
One could argue that PS4 Pro is beta testing "experimental" technologies for the PS5, just like PS3 (Cell SPUs) was beta testing for Compute Shaders... people may mock these technologies all they want, until they become mainstream enough.
Last but not least, don't forget that Moore's Law has slowed down quite a bit. Some people expected that FP64 would replace FP32 in consumer GPUs, but that would halve performance or it would require double the amount of transistors for the same performance. That's why FP16 and even INT8 are becoming a thing. Mobile GPU design is also another factor to consider.
~
In regards to Switch (Tegra X1), does anyone know if any Nintendo exclusives (Zelda BotW, MK8 Deluxe, Splatoon 2) utilize RPM/2xFP16? 3rd party devs are even less likely to use it, but who knows... it's an interesting feature nonetheless.
When it comes to PS4 exclusives, I haven't heard of ND or GG utilizing it in current games. Maybe in the future.
Remember when nVidia had released GeForce 256 (very expensive back then) and everyone said that T&L was a "useless gimmick"? Most people had 3Dfx graphics cards and thought that a fast enough CPU would suffice for vertex transformation. T&L/DX7-ready games took a while to become mainstream... and look where we are now.
People will always be skeptical when a paradigm shift disrupts the status quo.
Negotiator
Member
Not bad:
Dreamwriter
Member
Maybe Im missing something, but I just...dont get how this is a big feature. I just last week wrote a shader for pc using half-precision variables, its a very old feature that every single shader programmer should be using, you dont use higher precision than you need. Its not difficult or anything either, you just replace floats with half or fixed. And all hardware has supported it for years. Heck, heres a bit from the Unity 3.55 documentation, from 2012:
Admittedly, I havent been programming shaders very long, so I could very well be missing something.
Unity 3.55 said:
Admittedly, I havent been programming shaders very long, so I could very well be missing something.
I think the discussion here is the performance gain from fp16 and hardware designed with it.
Brute force isn't how I would word it given the DX12 customisations in the hardware. Feels more like they want to encapsulate backwards compatibility and ensure the newer consoles can work with previous games.
Sony, likewise have likely implemented their first party dev feedback straight into the design if the Pro.
The big deal is that instead of getting the same performance out of fp32 & fp16 you will get 2X the performance out of fp16 when it can be used.
Sad Affleck
Member
Image not showing up in your post but I re-linked it .
Indeed those are decent gains. I'm interested in seeing what can be achieved in actual games but this is very promising.
Dreamwriter
Member
Yes, but my point is that as far as I know all hardware based on modern programmable shader systems have fp16, from the original XBox onward, including cell phones, and all games should have been using it for years and getting better performance. We shouldnt be seeing any new performance gains from new games, because its not some amazing new technique made for 4k gaming that a game would need to be designed around, its a simple pixel-shader optimization feature that lets the GPU process half as much data per pixel, thus increasing the speed that the pixel will be rendered. Saying Wolfenstein II and Far Cry 5 will support it is like saying those games are going to support baked lighting, or even polygons.
Script of the Bridge
Banned
As soon as I read FP16 in the title I knew onQ123 would be here
As others have explained, you can process fp16 at double rates and int8 at quad rates. When previously if you used a fp16 variable it would only be executed at the same speed as fp32. Now, there is also a halfway point where you could stuff two fp16 variables into the same fp32 register but that would only relieve register pressure and not double execution speed.
So to be clear, currently you using fp16 variables your gpu isn't executing them any faster than fp32.
dr_rus
Member
You won't get 2X performance out of FP16 even when looking only at the shader using FP16. The slide you've shown is giving you the typical performance gains from using 16 bit math (note that INT16 and FP16 are different things, and they are supported differently by the h/w base) in specific shaders in a demo made specifically to showcase this feature.
Their final impact on the whole frame rendering time will be way less than stated there. I expect games to get less than 10% of performance above FP32 from double speed FP16 without loss of quality. Most graphics calculations pretty much require FP32, some graphics calculations which can live with FP16 are in fact not limited by math throughput. It's not a magic "2X" feature.
You're really overthinking things.
You will get 2X the performance out of your fp16 code as you would have without RPM I'm not talking about 2X the performance of the full game.
No you won't. Only if the shader in question is 100% alu bound.
You're getting 2X the theoretical performance.
why do you act like the same rules don't apply with the numbers we been getting for years?
killatopak
Member
Can anybody give real world example of how this helps in games? Like which games supported this feature and what were the benefits?
I'm really starting to like the cut of your jib in certain topics. Expect a pm later on some of the networking stuff we discussed.
I also remember people thinking powervr and 3dfx would lead things as well.
Polygonal spirte mentioned snake pass for switch which is a great example cause without it or optimization in general the game wouldn't run that well on the platform.
what?
Sad Affleck
Member
He really wants you to say that you will get 2x performance for console war reasons.
Whenever we have gotten specs for GPUs they have been the theoretical peak floating point performance number & that doubles for fp16 when the GPU has RPM/double rate fp16.
Dictator93
Member
Without full context of those numbers, that reads as percentage speed ups of singular parts from whole effects (say, a single step in generating volumetric lighting, which may have 5-8 steps).
Or does that mean the Bloom in its entirety is now XX% faster than previous?
This only started when the new consoles launched and people needed an easy to use number that indicated performance. Before this generation no one used flops for gaming performance comparisons in the GPU space because they are quite useless for these comparisons, especially comparisons between different architectures. Even within a single architecture performance might not scale with flops.
I mean just look at the Pro. Even in GPU limited scenarios it doesn't achieve a 2.3 increase in performance like you would expect based on flops.