Evil Supremacy
Member
There was a thread about the ps4 pro and fp16 a while back which had some informative posts and links, worth a read.
http://www.neogaf.com/showthread.php?t=1297459&page=1
-
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.
Inside the Scorpio Engine: the processor architecture deep dive
- Thread starter Space_nut
- Start date
There was a thread about the ps4 pro and fp16 a while back which had some informative posts and links, worth a read.
http://www.neogaf.com/showthread.php?t=1297459&page=1
Shpeshal Nick
aka Collingwood
You misunderstand.
I'm not referring the the Pro now. I'm saying that had Scorpio included the ability for FP16, it would have allowed it to be the base model longer to keep up with newer games.
I'm not talking about vs Pro. It's much stronger than the Pro. I'm aware of that.
Liabe Brave
Member
Yes, this is the answer we want. Unfortunately, none of us know what a realistic estimation would be. People in this thread and elsewhere have posited between 10% and 35%. As I've previously posted, this would make Pro's effective power somewhere between 4.6TF and 5.6TF. But these are all just guesses, and it's unclear which is most realistic.
"Keep up" is wrong, Scorpio will always be more powerful than PS4 Pro. Its lack of packed math, if anything, may hurt its ability to stay as far ahead.
EDIT - I see you meant something else. Yes, having RPM would give Scorpio even more headroom to adapt to demanding future games.
That quote is talking about PC games. Console games, which typically have to deal with tighter bandwidth and memory constraints, use FP16 more often. Mobile games, with even tighter envelopes, use it more often still. There will be a continuum of effectiveness.
Because it is irrelevant, at least in terms of results. Developers will only use FP16 where increased precision is unnecessary or its lack imperceptible. This lowers the scope of how often it can be used, but won't degrade the onscreen effects.
It's not a narrative, it's just logic. If Pro can consistently make better use of its GPU than Scorpio, by definition that brings it closer to Scorpio in terms of results. Not "catch up", not "surpass", but closer.
This is entirely too sensible but I figured I'd let you know at least someone read it!
The way more significant fact than Scorpio not having double packed math is that the baseline (Xbone) is lower. They will always have to design around Xbone as the bottom line, and converting an Xbone game into a 4k game is just going to look worse than a PS4 game at 4k.
If it were irrelevant then developers would be building entire (console) games with FP16. They aren't, because it's relevant.
If it were irrelevant then developers would be building entire (console) games with FP16. They aren't, because it's relevant.
Shil Pencer
Member
are people trying to downplay the power of Scorpio....let it go people and enjoy all gaming
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
Actually something not requiring good sub-pixel positioning at the vertex stage could be entirely written in fp16. For instance Puyo Puyo Tetris (or almost anything inherently 2D).
So is FP16 support the secret sauce in the Pro to close the gap now?
Edit:
Please, not this shit again.
Edit:
Please, not this shit again.
Not true
dr_rus
Member
Pro's biggest shortage compared to Scorpio is memory bandwidth, not processing power. It will run out of memory bandwidth before running out of flops if someone will try to run the same code on it as on Scorpio.
Shader calculation precision has almost nothing in common with external buffers precision. Saying that you don't need FP32 because you're still outputting an image to a 24 bit RGB buffer is all sorts of misleading. You actually do need FP32 for calculations if you want that final INT8 image to look great. Because "Of course if you use fp16 in a wrong way, you get banding and other artifacts." - point is that for the major amount of gaming math you will always end up with artifacts when using FP16. The part of math which can be downgraded to FP16 won't lead to big performance gains - I've heard figures anywhere from +5 to +50% depending on who you're asking and what they mean.
Liabe Brave
Member
I said it was irrelevant to results; that is, you wouldn't see messed-up/banded/dithered graphics, because FP16 just won't be used if it negatively impacts the game. Of course precision is therefore relevant to whether or not FP16 is viable for a specific task. But I assumed you weren't talking about that type of relevancy, because you said people were neglecting the impact of lower precision. But in terms of it limiting use cases, no one was neglecting that.
But FP16 packed math improves bandwidth usage versus unpacked FP16. I'm not sure what you're getting at here.
No Phil someone posted Scorpio vs Pro specs & left out one of the biggest difference in the PS4 Pro specs & I fixed it but that seems to upset people even though it's a fact.
JamesHolden
Banned
This is laughable. Thank you for providing me with laughter.
Correct me if I'm wrong, but PS4 doesn't have double packed math and its the base line for PS4 Pro. As a matter of fact Sony has said there will be no Pro exclusives. So what triple A dev is going to realize FP16 double rate to its potential if they have to use their PS4 versions as the starting point?
If your talking about PS5 or 'next gen' games, It would be able to keep up even less in that case than Pro currently will be able to for Scorpio.
These machines are what they are because they have to be made on a set time table with certain limitations, same for Pro and Scorpio.
they can't be futureproofed, its the nature of technology
Electret
Member
You're missing the very simple point that FP16 is sufficient to represent some data sets without causing a loss of information. It's not an all-or-nothing scenario. Not everything requires FP32 precision.
Shpeshal Nick
aka Collingwood
So reading that article more, it looks like Ryse at 4K/60 might be possible.
Good. Lordy. Fuck if they can.
Good. Lordy. Fuck if they can.
JamesHolden
Banned
So FP16 is so important for gaming, that Nvidia pruposely neutered FP16 caluclation in its GTX 1080 cards so as stop people from buying GTX1080's instead of their tesla line for compute heavy workloads.
Here is Anandtech with a detailed writeup on this last June
So FP16 is SO Important for gaming, that Nvidia purposely neutered its flagship gaming cards. FP16 has such little limited applications in console gaming, unless we are talking Mobile like the switch, that citing numbers like 25-30% boost for the base teraflop number on the PS4 PRO isn't just absurd, it should be laughed at. Haha.
Here is Anandtech with a detailed writeup on this last June
So FP16 is SO Important for gaming, that Nvidia purposely neutered its flagship gaming cards. FP16 has such little limited applications in console gaming, unless we are talking Mobile like the switch, that citing numbers like 25-30% boost for the base teraflop number on the PS4 PRO isn't just absurd, it should be laughed at. Haha.
RowdyReverb
Member
It's like 2013 all over again. First are the spec difference denials, then comes acceptance of the spec gap and questions about whether or not the visual difference actually matters, and then gradual acceptance that one box is clearly more capable than the other.
It will be ok
It will be ok
Risk Breaker
Member
The point is, it is a great feature if used properly, regardless of how powerful the hardware is doing FP32.
When devs make features to take advantage of the PS4P's packed math like is already the case in Frostbite, they can't do the same with Scorpio to get extra performance out of it.
Also, exclusives on PS4P will probably take advantage of it, and exclusives on Scorpio would do it too if it actually had the feature.
I personally thought it was a given, and it's kinda disappointing to know it's not the case. Simple, really.
You kinda answered your own question. Nvidia limited the compute capabilties of their GPUs and there was no hardware to take advantage of FP16 in the desktop area, ergo devs would not bother to use a feature that is not there.
Also, from anandtech:
If Nvidia put this into Pascal, I guess it's not so laughable, huh? Haha.
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
Ryse is a 1600x900@30hz title.
Tripolygon
Banned
You are going in with the baseless assumption that Scorpio can't make consistent use of its 6TF GPU. Consoles are not PC and developers can make better use of GPU than on PC. The changes in Pro are there to make better use of the 4.2TF GPU, does not make it perform better than or bring it closer to an arbitrary 6TF GPU. That is illogical. The fact remains that Scorpio is 6TF and Pro is 4.2TF, there is no getting closer, or ifs and buts as they share a similar architecture.
Zvonimir Boban
Banned
4K yeah but I can't see 60fps, especially when I don't think Ryse managed a locked 30fps on Xbone.
dr_rus
Member
No, it doesn't. It improves LDS/register pressure as in the amount of on chip memory used during shader execution but it has absolutely no relation to the external memory.
It might be a fact but a rather useless one not worth mentioning. It set's up misleading expectations.
What gave you that impression? I've read the whole thing and didn't come to that conclusion.
Why yes that may be possible, and while that article may be a long, good and informative read about many aspects of the Scorpio cpu/gpu. This thread has derailed this into a "conversation" about this one little statement...
Which could mean Sony has out smarted MS once again. Even with a year head start. MS is still going to come up short. Maybe not in the short term, but long term when all developers have converted all the engines, tools, and middleware to FP16 double rate. Can't wait for Sony's 1st party to unlocked those 8.4TFs. GOW with FP16 double rate should look incredible.
Can Crusher
Banned
Maybe fp16 doesnt matter, maybe it does. We will see when we get to compare gears to los2, forza to gt etc
vpance
Member
I think it'll likely head higher towards 50% as devs gain more familiarity in how to handle it. From what I've read 2xFP16 will see worth while usage as time goes on, especially for console devs.
Between TDP limitations and the expense of moving to new node shrinks, next gen can only benefit with these sorts of optimizations. Full FP32 calculations is a complete waste of GPU time and is actually overkill for pretty much all games.
Of course people are neglecting it; that's the entire root issue with the whole push to compare PS4 Pro's 8.4TFLOPS at 16bit precision vs Scorpio's 6TFLOPS at 32bit precision.
LukasTaves
Member
It halves the size of each number (and thus the bandwidth). The precision is reduced by a fator of over 65k. That's the size of the downside people are ignoring XD
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
Liabe Brave
Member
As previously explained, some console games are already using FP16 in some areas, because there are benefits to it even without double-rate computation. The Frostbite engine was using FP16 back on PS3, Decima used it at PS4 launch, Ubisoft are using it, etc.
I'm sorry, but this doesn't make any sense. Of course Scorpio is well-tuned to use as much of its 6TF as possible, and devs will find ever more efficient ways to do that. But Pro is not just a 4.2TF device. It's also (in theory) an 8.4TF FP16 device. Of course, it can't get anywhere near that theoretical performance with real-world code. But if it can be that device for any amount of time at all, that will raise its effective power versus Scorpio, which can't take advantage of this mode. It's very much like comparing two machines, where only one of them has a core clock turbo mode. This may only operate a fraction of the time, but that will affect the power gap between the devices.
I could easily have misunderstood, but I thought what you describe is the effect of using FP16 at all, regardless of rate. What then is the difference between using FP16 and using packed FP16?
I think you are fundamentally misunderstanding what people are saying. When they say "FP16 can only be used in specialized cases", that is them explicitly not neglecting the effect of precision differences. They are accounting for precision differences, which is why FP16 can't be used at all times.
LukasTaves
Member
Ok, my post wasn't entirely correct, but neither is just comparing the mantissa and forgetting the expoent, and the extra bits of the fp32 expoent also gives it way more decimal places specially for larger numbers.
But to be completely correct: going fp32 to fp16 reduces the range of the numbers you can represent by a factor of over 65k
Beerman462
Member
The point is, it is a great feature if used properly, regardless of how powerful the hardware is doing FP32.
When devs make features to take advantage of the PS4P's packed math like is already the case in Frostbite, they can't do the same with Scorpio to get extra performance out of it.
Also, exclusives on PS4P will probably take advantage of it, and exclusives on Scorpio would do it too if it actually had the feature.
I personally thought it was a given, and it's kinda disappointing to know it's not the case. Simple, really.
You kinda answered your own question. Nvidia limited the compute capabilties of their GPUs and there was no hardware to take advantage of FP16 in the desktop area, ergo devs would not bother to use a feature that is not there.
Also, from anandtech:
If Nvidia put this into Pascal, I guess it's not so laughable, huh? Haha.
Consumer Pascal cards like 1070 and 1080 don't have this.
oldergamer
Member
The armchair warriors are strong in this one. Fp 16 isnt going to make the kind of impact people are saying. How does this factor into visuals can be answered by b3d.
Liabe Brave
Member
This matters considerably less for functions whose value ranges between 0 and 1, which do occur in games. Hence why a portion of computation is amenable to lower precision (though only a portion).
Some developers at Beyond3D are very bullish on FP16 usage, and others are more skeptical. About all we can say for sure is that devs are already using it, that its presence and effectiveness will vary from title to title, and that it won't let Pro match up to Scorpio.
Proelite
Member
In 2005, MS releases a console that was $100-$200 cheaper than Sony a year earlier that ran third party ports across the board much better.
In 2013, Sony releases a console that was $100 cheaper than MS that ran third party ports across the board better.
In 2016, Sony releases a console that might be $100 cheaper than MS's 2017 console that can achieve parity with the use of Vega features down the line.
If the trend holds up...
In 2019, Sony's PS5 will be a $100 cheaper than MS's 2020 console, be more powerful on paper, and be more advanced while running third party ports much better.
In 2022, Sony's PS5 pro will be the only high end console released, MS chases casuals with a Switch rip-off after Switch hits 100 million units sold.
CassidyIzABeast
Member
Yes its talking about PC but the comment also applies to these mid-gen refreshes. Very few games are going to be doing major RPM FP16 optimizations to their engines when 1 out of their 5 targeted platforms don't take advantage of it. RPM FR16 is a cool feature for Pro but its advantages won't be fully exploited until next gen.
Self_Destructive
Member
Oh my...
It's not a secret. It's an undeniably useful capability that the PS4 Pro has, and the Scorpio does not. It doesn't magically make the PS4 Pro stronger than Scorpio (and no-one in this thread is suggesting that)... but it does help.
dragonelite
Member
This its kinda funny to see the similarities, FP16 vs Esram readwrite Bandwidth etc.
LukasTaves
Member
That's true, in the [-1,1] range fp16 does apparently offer enough precision. But in those cases, you are likely more bandwidth bound than math bound, so the double math rate wouldn't matter all that much. Using half bandwidth however, could be a nice gain on Pro which apparently is bandwidth starved.
Say what you want but MS is absolutely genius. They took "Gaf/Reddit" mainstream. They threw napalm on a simmering oil fire. All this chip deep dive architecture shared memory bus DCC jargon although valid is the new this vs that. I don't ask my Doctor how certain chemicals in my prescriptions bind to particular receptors. I just know it works. Can't wait till E3!
Right, and then they turn around and double the FP32 number to arrive at a FP16 number to compare 'peak performance', completely neglecting that you simply cannot compare them, because FP16 needs to be used on a case-by-case basis for different tasks.
Literally the only reason FP16 is being discussed ad nauseum is that people are pushing it as a secret sauce a la ESRAM's low latency, or the move engines, or DX12.
Remember how the person who brought up this whole FP32 vs FP16 thing in the first place was pushing a conspiracy theory that Scorpio's 6TFLOPS figure was already taking into account FP16 and that the console was only 3TFLOPS in FP32? Like, if it wasn't obvious what onQ's agenda was then I don't know how it isn't now.
Go on
LOL you don't really know what's going on do you?