• Hey Guest. Check out your NeoGAF Wrapped 2025 results here!

Mark Cerny on The Verge (GPU, elite controller, etc)

That was such a ridiculous comparison, made absolutely no sense. And literally nobody is making the argument that it would make a 30fps game to 60fps.

The bolded is false. FP16 has been used in graphics rendering for a long long time. Heck even games on PS3 used it. Seriously, its not a new thing at all, all that is new is PS4 Pro and future AMD GPUs support 2 FP16 operations per clock cycle. Mobile GPUs support it.

Exactly I believe this will help mobile applications more than AAA console/pc rendering. Mobile gpus need that extra bit of optimization to compensate the low power usage and raw processing. The Pro gpu won't perform better than a comparable 4TF pc gpu especially with a jaguar CPU. I'd like to see if GDC will delve into this
 
If you look at the practise you may also claim that it's not a 4,2 TF machine because the practical performance is not matching the theoretical output since the game code is hitting various bottlenecks.

It's in theory a 4,2 TF FP32 machine and a 8,4 TF FP16 machine.
Nobody should try to sell explicit performance simply because of that, it's a "stupid" number which only shows one part of the whole picture.
I don't get the arguments for or against the notion of the hardware specs.

I think most people understand that you never reach the theoretical numbers in real world situations. However it's a decent point of reference to give us an idea what to expect from hardware.

The PS4 and XBO surely aren't hitting their theoretical numbers but the results are reflective of the gap between the quoted specs of the two systems.
 
It's a damn shame about the elite controller but I've been waiting for some sort of confirmation like this to buy a scuf controller so at least there are other options.
 
Definitions of those are debatable. Presumably their caveat would be "4K on most new games, unlike the Pro", similarly without the, er, 'optimizations' of PSVR.

Calling a three teraflop SP GPU the most powerful console ever launched would be so demonstrably false it would be ridiculous.

when they released the video PS4 Pro didn't exist.



You see how this work now?

PS4 Pro wasn't announced so to MS it didn't exist & Scorpio isn't on the market so to Sony it doesn't exist.

pxeOsM5.jpg



https://www.youtube.com/watch?v=dSXpKhSLCYE
 
Problem with that is the pro specs leaked long before Scorpio reveal and we have information on Scorpio so it does exist. Let's see how fallout 4 looks on this 4.2tf 32fp pro and 8.4tf 32fp pc.

And the Xbox Scorpio was revealed long before this Sony ad was made & what does the last part of your comment have to do with my post?
 
Problem with that is the pro specs leaked long before Scorpio reveal and we have information on Scorpio so it does exist. Let's see how fallout 4 looks on this 4.2tf 32fp pro and 8.4tf 32fp pc.

Yup, MS knew about the Pro before the official announcement and I wouldn't be surprised if Sony knew about Scorpio before it was officially announced as well.
 
Yup, MS knew about the Pro before the official announcement and I wouldn't be surprised if Sony knew about Scorpio before it was officially announced as well.


4TF , 6TF ,8TF ,12TF , 8TF + super duper special sauce , 12TF + The Power Of The Clouds to infinity!

giphy.gif
 
With that reasoning, it's a 16.8TF FP8 machine. Or 33.6TF FP4 machine.

The bizarre thing about this thread is that almost everyone wants to argue against onQ123, as if he's Sony's version of MisterXMedia. Yet, if anything, more tripe is coming out of his detractors than him.

The PS4 Pro, given what has been said, does have an 8.4TF GPU if the f is taken to refer to half precision floating point numbers (i.e. FP16). This means it it can theoretically perform 8.4 trillion FP16 operations per second. This isn't the case because, inherently, 1 FP32 operation takes twice the time of 1 FP16 operation which takes half the time of 1 FP8 operation, it is the case because the machine has hardware which can perform two FP16 calculations at the same speed as one FP32 calculation (which is not the same as being able to do one FP16 at twice the speed as one FP32 calculation!)

You couldn't have made the same claim about the PS4. You couldn't say it's GPU was capable of 3.68 trillion FP16 operations per second because it wasn't. There is a specific upgrade in the PS4 Pro's GPU which means the claim is legitimate.

The PS4 Pro's GPU is undeniably capable of 8.4 trillion FP16 operations per second. I don't think it is advisable to call it an 8.4TF machine because it is likely to confuse: floats are, by default, 32 bit. At the same time, it would be wrong to neglect the innovation in this GPU. It is significant, and though special sauce is a loaded term, it kind of is special sauce, albeit special sauce which will be used in many GPUs. I wouldn't be surprised if it becomes a feature in all GPUs in the future.
 
What it comes down to is what MS is referring to with regards to their 6TF bullet point. It's insane to think that it's anything other than full FP32 performance. Onq thinks he's got a reason to believe that it could be different.. and he might. But as I said before.. I'm SURE he knows himself that it's FP32, and not 16/32. He's just leaving the door open in case something absolutely insane happens, then he can say 'look.. I told you'. NOBODY is arguing the numbers. (theoretically) PS4 Pro is a 8.4TF FP16 machine.
 
The bizarre thing about this thread is that almost everyone wants to argue against onQ123, as if he's Sony's version of MisterXMedia. Yet, if anything, more tripe is coming out of his detractors than him.

The PS4 Pro, given what has been said, does have an 8.4TF GPU if the f is taken to refer to half precision floating point numbers (i.e. FP16). This means it it can theoretically perform 8.4 trillion FP16 operations per second. This isn't the case because, inherently, 1 FP32 operation takes twice the time of 1 FP16 operation which takes half the time of 1 FP8 operation, it is the case because the machine has hardware which can perform two FP16 calculations at the same speed as one FP32 calculation (which is not the same as being able to do one FP16 at twice the speed as one FP32 calculation!)

You couldn't have made the same claim about the PS4. You couldn't say it's GPU was capable of 3.68 trillion FP16 operations per second because it wasn't. There is a specific upgrade in the PS4 Pro's GPU which means the claim is legitimate.

The PS4 Pro's GPU is undeniably capable of 8.4 trillion FP16 operations per second. I don't think it is advisable to call it an 8.4TF machine because it is likely to confuse: floats are, by default, 32 bit. At the same time, it would be wrong to neglect the innovation in this GPU. It is significant, and though special sauce is a loaded term, it kind of is special sauce, albeit special sauce which will be used in many GPUs. I wouldn't be surprised if it becomes a feature in all GPUs in the future.
I don't think anyone is arguing the simple fact that it is capable of 8.4tf when using half floats...BUT, it's frankly an irrelevant number...it's impossible to code a game entirely in FP16...meaning that the Pro is completely incapable of using anywhere near that kind of power for gaming
 
I don't think anyone is arguing the simple fact that it is capable of 8.4tf when using half floats...BUT, it's frankly an irrelevant number...it's impossible to code a game entirely in FP16...meaning that the Pro is completely incapable of using anywhere near that kind of power for gaming

Entire PS3 games were made using FP16, same with mobile.

FP16 will only see more and more usage down the line, no doubt.
 
Entire PS3 games were made using FP16, same with mobile.

FP16 will only see more and more usage down the line, no doubt.

Which PS3 games? (Genuinely curious)...and are you really suggesting we're going to get 8.4tf mobile games on the PS4 Pro? I mean, seriously?
 
They're OK with offering a premium console for enthusiasts but not OK with doing the same with controllers?

Makes zero sense.

Well I can understand it in terms of online play. It's a variation on the whole allowing m/kb on consoles argument and how that would supposedly upset the balance.
 
Which PS3 games? (Genuinely curious)...and are you really suggesting we're going to get 8.4tf mobile games on the PS4 Pro? I mean, seriously?

Well, all of them I guess. RSX was only capable of FP16. Point is, games can still look good in FP16, and is apparently still viable in much of today's games.

What about XBOX 360? Don't last-gen consoles support FP32?

360 had up to FP10 support.
 
Well, all of them I guess. RSX was only capable of FP16. Point is, games can still look good in FP16, and is apparently still viable in much of today's games.

I'm pretty sure that's wrong. :/
This is the first link on Google: http://www.anandtech.com/show/1689/3

NVIDIA confirmed that the RSX is features full FP32 support, like the current generation GeForce 6 as well as ATI's Xbox 360 GPU.

But even beyond that, based on what I've been hearing since forever and as far as I understand things, I'm pretty sure that's not right.
I think the SPEs were only FP16 though.

Also, I don't think any graphically intensive mobile games can pull that off either. Link?
 
I'm pretty sure that's wrong. :/
This is the first link on Google: http://www.anandtech.com/show/1689/3



But even beyond that, based on what I've been hearing since forever, I'm pretty sure that's note right.
I think the SPEs were only FP16 though.

Right, but still, it was rarely if ever used since it was performance heavy. On PS3 you didn't have native AA support with FP16/32 render targets so a lot of devs opted to use RGB8 with clever math tricks to get HDR in their games.
 
I think the SPEs were only FP16 though.

The Cell SPU instruction set supported both single- and double-precision floating point data types and doesn't have any half-precision support at all.

There's no question that having FP16 available is a boon for developers who are willing to code specifically for it. The range and precision for these values is incredibly limited, but there are definitely algorithms where it makes sense – not just for throughput but also for bandwidth savings. It's just not a panacea. You can't simply assume that you'll get double the throughput across a wide range of scenarios.

As I and others have stated, half-precision isn't something new. It has been in use for quite some time on other architectures, it's just new to AMD's GCN. Part of the challenge here will be that PS4 doesn't support the data format, so shaders need to be rewritten specifically for the PS4 Pro to see any advantages.
 
The bizarre thing about this thread is that almost everyone wants to argue against onQ123, as if he's Sony's version of MisterXMedia. Yet, if anything, more tripe is coming out of his detractors than him.

The PS4 Pro, given what has been said, does have an 8.4TF GPU if the f is taken to refer to half precision floating point numbers (i.e. FP16). This means it it can theoretically perform 8.4 trillion FP16 operations per second. This isn't the case because, inherently, 1 FP32 operation takes twice the time of 1 FP16 operation which takes half the time of 1 FP8 operation, it is the case because the machine has hardware which can perform two FP16 calculations at the same speed as one FP32 calculation (which is not the same as being able to do one FP16 at twice the speed as one FP32 calculation!)

You couldn't have made the same claim about the PS4. You couldn't say it's GPU was capable of 3.68 trillion FP16 operations per second because it wasn't. There is a specific upgrade in the PS4 Pro's GPU which means the claim is legitimate.

The PS4 Pro's GPU is undeniably capable of 8.4 trillion FP16 operations per second. I don't think it is advisable to call it an 8.4TF machine because it is likely to confuse: floats are, by default, 32 bit. At the same time, it would be wrong to neglect the innovation in this GPU. It is significant, and though special sauce is a loaded term, it kind of is special sauce, albeit special sauce which will be used in many GPUs. I wouldn't be surprised if it becomes a feature in all GPUs in the future.

That's mathematics son they can argue with me but they can't argue with figures!


K29_KedAHMs8dCemN07cu2xRhX4txrmTadJndkmfRSE8TLfUr4DCe6PYPZJhYPGGxKPLxZ7foElFLbv5z_HawzKrOb4=w426-h240-n
 
You couldn't have made the same claim about the PS4. You couldn't say it's GPU was capable of 3.68 trillion FP16 operations per second because it wasn't. There is a specific upgrade in the PS4 Pro's GPU which means the claim is legitimate.
Indeed.
If I remember correctly Ps4 GPU cannot calculate using fp16 precision.
It can store fp16 values in registers, but math is done in fp32.
 
Indeed.
If I remember correctly Ps4 GPU cannot calculate using fp16 precision.
It can store fp16 values in registers, but math is done in fp32.
The PS4 can't save register space with FP16, this ability came first with GCN Gen 3, the PS4 uses GCN Gen 2 (at least as a basis).
If the PS4 could store native FP16 values then devlopers would already using it, since it would save registers and increase practical performance.
 
The PS4 can't save register space with FP16, this ability came first with GCN Gen 3, the PS4 uses GCN Gen 2 (at least as a basis).
If the PS4 could store native FP16 values then devlopers would already using it, since it would save registers and increase practical performance.
Pretty sure this is what was talked about in this presentation. (Packing 2 16bit values in 32bit register)


https://www.google.fi/amp/s/michald...ions-for-gcn-digital-dragons-2014-slides/amp/

https://www.youtube.com/watch?v=Bmy3Tt3Ottc
 
That's about special packing ops for multiple render targets, not for average shader code itself.
(Minute 21:40 if somebody else wants to check it out)
 

Right, and the post you've quoted is from nine years ago. HDR is going to be expected on any AAA title going forward, and vector math at low precision is going to have obvious artifacting implications when used for lighting except under fairly controlled conditions. There will be plenty of FP16 opportunities, but showcase titles aren't going to get anything like a 2-1 advantage over the original FP32 shaders for PS4.

It's definitely good, but it overselling the benefits sets unrealistic expectations.
 
Right, and the post you've quoted is from nine years ago. HDR is going to be expected on any AAA title going forward, and vector math at low precision is going to have obvious artifacting implications when used for lighting except under fairly controlled conditions. There will be plenty of FP16 opportunities, but showcase titles aren't going to get anything like a 2-1 advantage over the original FP32 shaders for PS4.

It's definitely good, but it overselling the benefits sets unrealistic expectations.

That was posted today
 
It's a damn shame about the elite controller but I've been waiting for some sort of confirmation like this to buy a scuf controller so at least there are other options.

Tell me about it.

I want a SCUF but do I want to pay that price? Hoping for a sale between now and Christmas.

The trade in thing with them is good, although still kinda expensive.
 
Tell me about it.

I want a SCUF but do I want to pay that price? Hoping for a sale between now and Christmas.

The trade in thing with them is good, although still kinda expensive.

I have a scuf and can only say: Don't buy one! The quality sucks.

After receiving mine, the grip had sharp edges that were really annoying. After repairing I noticed that the sticks are not perfectly centered. The camera accelerates faster when I push the stick up in comparison to pushing it down. My friends scuf was completly broken. Input lag and broken sticks.

Save the money and don't buy one.
 
D'oh! Thanks for calling that out – evidently watching a football game and posting means less focus than I expected. The point stands, that HDR is pretty much an expectation for PS4 Pro titles. FP16 is extraordinarily common for mobile titles, but it will find somewhat narrower applications on high-end consoles.

Yeah, don't read too much into that.
 
D'oh! Thanks for calling that out – evidently watching a football game and posting means less focus than I expected. The point stands, that HDR is pretty much an expectation for PS4 Pro titles. FP16 is extraordinarily common for mobile titles, but it will find somewhat narrower applications on high-end consoles.

Just a guess but I think that the Work distributor , ID Buffer & other hardware is going to lighten the load so that FP16 can be used a lot more, we still don't know the actual hardware that is inside of the PS4 Pro besides mobile games are built around having slow memory also so games on a console using FP16 does not = mobile game tech limited.


It could have 64 tlfops blast processing for all l care no UHD BluRay no buy.

But it has a Blitter Mega Processor! are you no longer part of The Amiga Brotherhood?


 
Just a guess but I think that the Work distributor , ID Buffer & other hardware is going to lighten the load so that FP16 can be used a lot more

The specific features named are absolutely unrelated to the tradeoffs between single- and half-precision representations. Better work distribution just means better utilization of functional units, which could improve shader throughput independent of whether you're manipulating half- or single-precision values (or it could just enable efficient scheduling of half-precision work in the first place, chances are we won't get that level of detail confirmed.) The ID buffer enables specific algorithms that have little or nothing to do with the precision and range of data being manipulated in traditional vector and fragment shaders. It's there specifically for reconciling content rendered in successive frames.

Sure, it's possible that there's some other custom hardware at work. Holding out hope for something miraculous is far from wise, but it'll be fascinating to see what details we get. I keep expecting to hear about dedicated hardware to enable checkerboard reconciliation at scanout to avoid a redundant write/read bandwidth costs for the final frame buffer but there's little to suggest that it's actually the case other than the claim that checkerboard rendering is essentially free.

Watching new console launches is exciting. And exhausting.
 
The specific features named are absolutely unrelated to the tradeoffs between single- and half-precision representations. Better work distribution just means better utilization of functional units, which could improve shader throughput independent of whether you're manipulating half- or single-precision values (or it could just enable efficient scheduling of half-precision work in the first place, chances are we won't get that level of detail confirmed.) The ID buffer enables specific algorithms that have little or nothing to do with the precision and range of data being manipulated in traditional vector and fragment shaders. It's there specifically for reconciling content rendered in successive frames.

Sure, it's possible that there's some other custom hardware at work. Holding out hope for something miraculous is far from wise, but it'll be fascinating to see what details we get. I keep expecting to hear about dedicated hardware to enable checkerboard reconciliation at scanout to avoid a redundant write/read bandwidth costs for the final frame buffer but there's little to suggest that it's actually the case other than the claim that checkerboard rendering is essentially free.

Watching new console launches is exciting. And exhausting.


I think the fact that they are able to track the exact location of polygons without any help from the GPU processing power should open up doors & not just for checkerboard rendering & AA.

I think right now is more about 4K output so that's what is being talked about but I can't picture devs having access to new tools & data & not finding new ways to do things to take advantage of what they have in a closed platform.
 
Well Hello people


It'll also be challenging to see that the Pro features 64 ROPs, up against Xbox One X's 32.

Leaked documentation from the Sony SDK confirms this, and likely it's a factor of the platform holder adopting a 'mirror image' strategy on the GPU to ensure compatibility with the base system - the 32 ROPs of the original PS4 is doubled up along with everything else. As things stand, doubling up on ROPs is an intriguing postscript to what we know about the established Pro spec, but the docs also confirm that it's theoretically impossible to fully utilise them - the memory bandwidth just isn't there.
http://www.eurogamer.net/articles/digitalfoundry-2018-ps4-pro-and-xbox-one-x-processors-compared

 
I use a Nacon so I have my "elite" controller. And before someone complains about its d-pad in comparison to the damn good standard DS3 and DS4 d-pad ... I use the hori fightpad in that instance.
 
Last edited:
Top Bottom