• 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.

AMD Vega Features Leaked

I

icecold1983

Member
michael92 said:
Really bad for gaming though and any 3D spatial application for that matter. Whenever I see gaming related FP16 I cringe inside.

FP16: 2,048 integers
FP32: 16,777,216 integers
FP64: 9,007,199,254,740,992 integers

I don't like the regressing trends of mixed precision and the information loss is definitely not worth the pathetic 2x jump in performance. Addressing complex and compelling worlds, particularly in VR is going to require 64bit computing.

With 32bits you could represent a 16.7 kilometer grid with mm accuracy.
With 64bits you could represent a 9 million kilometer grid with micrometer accuracy.
Click to expand...

Postprcessing takesup a non trivial part of every frame rendered. There various other parts of a frame that just dont need precision over 16bit
 
SenjutsuSage

SenjutsuSage

Banned
AmyS said:
Yeah, I agree. I mean even the PS4 Pro GPU, while largely based on Polaris, is using some features beyond Polaris (Mark Cerny said). So Polaris GPUs for PC cards were new as of mid 2016, and PS4 Pro launched just months later. Why would Microsoft & AMD not do something similar for 2017 with Vega architecture in Scorpio? There's little chance their not, or as you put it, 100% it's going to be Vega.

Just because Scorpio is almost certainly going to be Vega does not mean it has to be equivalent to Vega 10 with HBM2. We're looking at 320+ GB/sec which could be handled fine with a 384-bit bus connected to 12GB GDDR5.
Click to expand...

Yea, no way it will equal Vega 10. They are shooting for a particular performance target and power envelope, and will be satisfied with hitting that. I do suspect, however, they will be exceeding that 6Tflops number, simply because it might come easy, and without very much effort. A lot of people won't agree with this, but I think Microsoft played it safe, intentionally undershooting where they suspected they could end up with Scorpio from a performance perspective.
 
I

icecold1983

Member
michael92 said:
I see a larger trend for regressing to FP16, not just for post. Like that dreadful racing game on the PS4 Pro. HDR/ 10-12bit color is going to make FP16 post insufficient anyway.
Click to expand...

The top down view racer covered by DF? Were there visual issues stemming from fp16 usage?
 
I

icecold1983

Member
michael92 said:
Aside from it looking like an upscaled PS2 game? Exponential decrease in information for a mere performance doubling is nuts in my opinion. The slowdown of Moore's law is making devs lunatic.
Click to expand...

Isnt it identical to the ps4 version minus 2x msaa and plus 4x the resolution? Base ps4 doesnt have double rate fp16
 
L

LukasTaves

Member
icecold1983 said:
Works just fine for post processing and various other algorithms
Click to expand...
I can't imagine it being good for anything graphic related really.

If we talk about space it's got terrible precision when not around 0. If we are talking about pixels it's just not enough colors.

I could see it being used for highmaps, perhaps even normalmaps if you are really memory/bandwidth bound.

But there's a reason why most modern implementations of this format used it only as a store/load format and converted to fp32 for doing any calculations.
 
R

Rushersauce

Banned
michael92 said:
Aside from it looking like an upscaled PS2 game? Exponential decrease in information for a mere performance doubling is nuts in my opinion. The slowdown of Moore's law is making devs lunatic.
Click to expand...
Man, both Nvidia and AMD stopped pushing FP64. I don't know much about it, but why?
 
I

icecold1983

Member
vpance said:
From what I've read, even FP32 is overkill for a lot of stuff. FP16 + FP32 is the way forward.
Click to expand...

Also what ive read from various developers/graphics guys. Lead graphics engineer of trials games, lead engineer of the order, Johan of dice, Andrew formerly of intel etc

Fp32 becoming standard was more of a political thing
 
S

Skux

Member
Honestly this is nothing but jargon to me and a lot of people. At the end of the day the only thing people will care about is frames per dollar.
 
A

AmyS

Member
SenjutsuSage said:
Yea, no way it will equal Vega 10. They are shooting for a particular performance target and power envelope, and will be satisfied with hitting that. I do suspect, however, they will be exceeding that 6Tflops number, simply because it might come easy, and without very much effort. A lot of people won't agree with this, but I think Microsoft played it safe, intentionally undershooting where they suspected they could end up with Scorpio from a performance perspective.
Click to expand...

Yep. Wouldn't be surprised to see somewhere around 6.4 to 6.8 TF without any big changes to cooling, etc.

i.e just 6.55 TF would be 5X the performance of OG Xbox One. Otherwise we're still looking at a 4.5X increase with the straight 6 TF. Either way, it's gonna be nice for devs & gamers alike.
 
D

dogen

Member
LukasTaves said:
I can't imagine it being good for anything graphic related really.

If we talk about space it's got terrible precision when not around 0. If we are talking about pixels it's just not enough colors.

I could see it being used for highmaps, perhaps even normalmaps if you are really memory/bandwidth bound.

But there's a reason why most modern implementations of this format used it only as a store/load format and converted to fp32 for doing any calculations.
Click to expand...

Sebbbi's talked about it having a lot of uses on b3d. Iirc, he also said INT16 can replace INT32 in many cases.
 
T

tuxfool

Banned
michael92 said:
FP16 is the easy way out when you hit a wall of node shrinks. Here is our totally 20Tflops card, please buy it for 5000$.

FP32 open world games are going to hit hard map size limits soon (16.7x16.7km with a 1mm step) and there probably won't be a gaming 64bit GPU available by then. Truly sad.
Click to expand...
Inverse logarithmic depth buffers work fine for world scale rendering. You don't need to use anything higher than fp32 if you use camera relative rendering. If you need fp64 coordinates you just need to do that on the CPU, the GPU only needs to handle fp32 then translated to the camera position.
 
T

tuxfool

Banned
michael92 said:
Does that render without artifacts when viewing very far distances (say a view from the sky) or are you still limited by the rendering distance of FP32? If it is the latter, then my point still stands and FP64 rendering is going to become necessary in gaming.
Click to expand...
I suggest you look at Outerra and tell me if that matters for the forseeable future. Or look at what Star Citizen is doing with its planets.
 
L

LukasTaves

Member
onQ123 said:
That might be the way you see things but the truth is that it's peak performance is 8.4tf
Click to expand...
Yeah, it's peak performance on a meaningless metric that is pretty much useless.

But if that's the case, there's no way they will release a console that has less than 12tf on that same metric.
 
P

pottuvoi

Banned
michael92 said:
FP32 is certainly not overkill in gaming. The problem with floating point math is that it loses precision the more you move away from zero. By the time you reach 16.7 million, you can't even address a single decimal.
Click to expand...
For things that do not need precision, it is overkill.

I would vager that many vertex operations would work fine on fp16, like culling or small change in topology.
Also things like ssao really do not need precision.

In ALU limited cases there should be nice optimization possibilities.
 
T

tuxfool

Banned
michael92 said:
When dealing with really large objects like planets, it obviously should be really subtle, but if you are looking at a skyline of a detailed world with buildings, artifacts would be far more noticeable.

Thread in question.
Click to expand...
That is a bug that isn't present, it was just a camera bug that has since been corrected.

Detailed objects such as buildings would be mesh merged with other objects or simply culled just as they are now. Having a high polycount object on the horizon would just be wasteful.
 
Locuza

Locuza

Member
michael92 said:
[...]Exponential decrease in information for a mere performance doubling is nuts in my opinion. The slowdown of Moore's law is making devs lunatic.
Click to expand...
There is nothing lunatic about it, there are just reasonable decisions and compromises.
Most developers will not use FP16 for stages and effects where you would see terrible image quality loss, in most cases we should see FP16 where the image degradation is very small next to not noticeable at all.

It's a fine feature to have and I'm looking forward to see how developers will author their shaders in FP16 and what the performance impact will be.
 
T

TheGreyHulk

Member
Deeke[VRZ];227552275 said:
By the way, if Project Scorpio does use Vega, it'll be Vega 11 not Vega 10.

Vega 10 will be the most powerful cards that hit the highest TFLOP performance and draw the most power. Just like Polaris 10 is used in the most powerful RX 400 (the 480) series GPU.

Vega 11 will also likely use GDDR5/X VRAM instead of HBM2 to reduce costs.

Scorpio with Vega 10 an HBM2 would skyrocket costs to like $1200 and it'd need to basically be a PC to operate.

Sure they could scale everything, but don't count on Scorpio meeting these specs at all--microsoft has already confirmed 6TFLOp, which is actually closer to Polaris 10 Ellesmere than Vega.
Click to expand...
Ummmm... that's not how buying a custom chip from AMD works... Vega 10/11 are just names for the chips AMD is making for their GPUs they plan on selling to consumers.

This has NOTHING to do with custom chips, such as the one MS is buying for scorpio. I fully believe MS is using Vega in their custom APU from AMD because it is how they'll be able to pack 6TF on the chip with the cpu cores and not have a gigantic chip since vega is more efficient than polaris.

Vega is just the architecture. MS works with AMD to design a chip for them and they'll put x amount of compute units and x amount of cpu units and they'll manufacture it for MS on a single chip.
 
Caayn

Caayn

Member
Head.spawn said:
How many maths does it take to run Battlefront DLC though?
Click to expand...
1PFLOPS FP128 duh
bonk.gif

Agent_4Seven said:
Oh my... They better release all Vega cards with water cooling.
Click to expand...
As long as the cooling is good, you can count on AIB for this, I don't have a problem with larger power consumption. A TitanXP draws north of 250w.
 
MightyKAC

MightyKAC

Member
iHaunter said:
If it competes with a 1080Ti at a better price-point I'm for it.
Click to expand...

Pretty much this.

I readily admit to understanding about 0.5% the stuff in the OP but what I DO know is that NVidia will have to respond to it and the competition that will result from it will be nothing but good news for the end user.
 
H

hodgy100

Member
@michael92's general posting theme of floating point precision and world size.

I dont think you quite know what you are talking about FP32 is more than enough for our co-ordinates system in games you go up to FP64 and all of a sudden you take a 2* perf hit because your math takes twice as long to compute because you couldn't think of a better solution. Bumping up to FP64 to combat loss of precision when you reach the numerical limits of fp32 is a brute force solution you just center everything around 0,0 and move the root nodes of your world chunks (because you would want to load an open world in chunks) we already have software that map the entire world this is a solvable problem.

like i can't name anything in a game that will require 64 bits of precision games are supposed to be flawed approximations which is why game physics sometimes screw up, or you get the odd graphical corruption, these approximations are done so that you can get actual realtime performance. FP64 will be for scientific research & computing purposes.

Star citizen's issue with floating point precision sounds like a bad engineering decisions. they knew they were going to be dealing with large distances so they should have engineered around that.
 
D

Durante

Member
How, exactly, would you "engineer around" the issue Star Citizen faces (of representing mm-level distances and lightyear-level distances in the same positional coordinate space)?

I mean, you could answer something like "separate coordinate systems", but merging and splitting those on-demand while keeping all of them in sync seems complex enough that it is by no means certain to be a better engineering solution than simply going with doubles.

Of course, all of that is rather beside the point discussed in this thread, since you wouldn't use those FP64 coordinates for rendering.
 
O

onQ123

Member
LukasTaves said:
Yeah, it's peak performance on a meaningless metric that is pretty much useless.

But if that's the case, there's no way they will release a console that has less than 12tf on that same metric.
Click to expand...

It's time to wake up, FP16 is not useless a lot of you are just stuck in your ways & don't want to accept that things are not as you thought.
 
S

Sir Abacus

Member
Durante said:
How, exactly, would you "engineer around" the issue Star Citizen faces (of representing mm-level distances and lightyear-level distances in the same positional coordinate space)?
Click to expand...

I'm guessing because when you're working at lightyear level distances the absolute precision doesn't really matter for putting the dot of a star on the screen. Single precision floats can go to a maximum of 3.402823 × 10^38. 3.4 x 10^38mm works out to be 10^19 light years to play with. You can represent the whole universe and then some just by using the highest precision at the mm scale.
 
L

LukasTaves

Member
dogen said:
Sebbbi's talked about it having a lot of uses on b3d. Iirc, he also said INT16 can replace INT32 in many cases.
Click to expand...

I looked up on some of his posts. Turns out that indeed, according to him, many post process are done with normalized vectors that falls into the [-1,1] range and for that fp16 would be suited.

He also mentions that it could be used for a LDR pass (some HDR implementations like Halo 3 mix a LDR and a MDR pass for delivering a HDR image).

For spatial math it's no good, unless you use many local coordinates to always work very locally to the coordinate center...


So I'm changing my tone, fp16 is not completely useless. There's indeed some cases were it could be applicable, though I still wonder it's worth the trouble, because mixing fp16 and fp32 math seems really boring to get everything right, specially if you don't know exactly the magnitude of every variable... Specially when in the aforementioned use cases are more bandwidth bound than math anyway (though there are some bandwidth savings to be made as well).

I think if fp16 has indeed any merit we will see it first on mobile, they will probably push it there as alu time and bandwidth matters a lot due battery life concerns, and from there make into console and pc space.

michael92 said:
I see a larger trend for regressing to FP16, not just for post. Like that dreadful racing game on the PS4 Pro. HDR/ 10-12bit color is going to make FP16 post insufficient anyway.
Click to expand...

Does that game actually use any fp16 shader or it was just speculation on DF part?

I honestly see no reason as to why they would have to resort to it, it's a very simple looking game, with very simple shaders, that a regular ps4 would have no trouble going over 1080p, and they even dropped the AA when running 4k on pro.

onQ123 said:
It's time to wake up, FP16 is not useless a lot of you are just stuck in your ways & don't want to accept that things are not as you thought.
Click to expand...

In the very least it has a very limited usability due the incredible low precision. Fp16/int16 has an okay range if that's what the problem needs, but very low precision. Really, for any number that doesn't start with 0.something.

But, there's indeed some uses that doesn't fall over that range. however it's still, a very limited portion of the math in a frame, that's not even ALU limited in the first place, so I doubt you are going to see any meaningful gain unless you are completely ALU bound and missing your target framerate by a few ms.
 
I

icecold1983

Member
LukasTaves said:
I looked up on some of his posts. Turns out that indeed, according to him, many post process are done with normalized vectors that falls into the [-1,1] range and for that fp16 would be suited.

He also mentions that it could be used for a LDR pass (some HDR implementations like Halo 3 mix a LDR and a MDR pass for delivering a HDR image).

For spatial math it's no good, unless you use many local coordinates to always work very locally to the coordinate center...


So I'm changing my tone, fp16 is not completely useless. There's indeed some cases were it could be applicable, though I still wonder it's worth the trouble, because mixing fp16 and fp32 math seems really boring to get everything right, specially if you don't know exactly the magnitude of every variable... Specially when in the aforementioned use cases are more bandwidth bound than math anyway (though there are some bandwidth savings to be made as well).

I think if fp16 has indeed any merit we will see it first on mobile, they will probably push it there as alu time and bandwidth matters a lot due battery life concerns, and from there make into console and pc space.



Does that game actually use any fp16 shader or it was just speculation on DF part?

I honestly see no reason as to why they would have to resort to it, it's a very simple looking game, with very simple shaders, that a regular ps4 would have no trouble going over 1080p, and they even dropped the AA when running 4k on pro.



In the very least it has a very limited usability due the incredible low precision. Fp16/int16 has an okay range if that's what the problem needs, but very low precision. Really, for any number that doesn't start with 0.something.

But, there's indeed some uses that doesn't fall over that range. however it's still, a very limited portion of the math in a frame, that's not even ALU limited in the first place, so I doubt you are going to see any meaningful gain unless you are completely ALU bound and missing your target framerate by a few ms.
Click to expand...

its not NEARLY going to double general performance like some warriors like to throw around, but it will be a real optimization use case that the better developers will find smart uses for.
 
You must log in or register to reply here.

Similar threads

AMD "Strix Halo" Zen 5 Mobile Processor Pictured
Rumor Analysis
38
2K
DaGwaphics replied
AMD Announced AMD Radeon PRO W7000 with up to 48GB of VRAM
News Hardware
34
2K
Hudo replied
  • Poll
PLAYSTATION 6: Potential Innovations, Features, Business Strategies & More (Speculation)
Opinion Analysis Platform 2 3 4
181
13K
hussar16 replied
AMD reports an increase in revenue by 44% YoY
Business Hardware
11
1K
GHG replied
Next-Gen NVIDIA GeForce RTX 4090 With Top AD102 GPU Could Be The First Gaming Graphics Card To Break Past 100 TFLOPs
5 6 7 8 9
416
36K
Black_Stride replied
Top Bottom