Playstation 5 [PS5] [Release November 12 2020]

Come on now, don't make me have to defend Sony PlayStation 5 again... You know I hate that! ;)

Let's not turn this into meme-central.
 
I've said this before on here, but there continues to be a lot of confusion regarding the dynamic clock situation. I think Sony made a good choice by implementing it, as it smartly improves performance in all scenarios within their power and heat envelopes. However, while I have no issue with stating its maximum clocks, advertising it as a 10.28TFLOP GPU is pretty disingenuous since it's technically incapable of such performance.

I've heard mention here that maybe it can do it, but potentially only for short periods or something. That's also not true.

The PS5 leverages a dynamic clock to reduce power consumption as more of its overall compute resources are leveraged. Or as Cerny put's it, the more "demanding" the game is. It's only capable of max clocks when at some unknown percentage of ALU usage, not 100%. So it's incorrect to multiply it's entire ALU capacity at max clocks since that isn't something it can or would do. In order to determine the actual max TFLOPs number, we'd need to know the clock speed it's capable of sustaining with ALL its compute active.
 
In the Mark Cerny presentation he said "constant power" approach can push GPU clock to 2.23GHz while "constant frequency" approach will
have trouble to maintain 2GHz.

Is there any reason why it happens? If the cooling solution is capable of 2.23GHz under heavy workload, why will there be problem to use
"constant 2GHz" even there is some power spike?
 
In the Mark Cerny presentation he said "constant power" approach can push GPU clock to 2.23GHz while "constant frequency" approach will
have trouble to maintain 2GHz.

Is there any reason why it happens? If the cooling solution is capable of 2.23GHz under heavy workload, why will there be problem to use
"constant 2GHz" even there is some power spike?

In the PS5 equation, power is fixed. So there are 2 variables to contributing to that power draw. 1) The amount of active compute and 2) frequency. So as you increase active compute, you decrease frequency and power remains constant. As you decrease active compute, you can increase frequency. This fact has nothing to do with the cooling solution or heat since the maximum parameters defined for those are fixed.

And actually that quote is pretty informative. What he's saying is that designing for peak loads, or in his terms "worst-case scenario", and fixing clocks to that you'd end up with a GPU struggling to hit 2GHz. So If we use Cerny's own number then the PS5 should be rated at 9.25TFLOPs. Interesting number no?
 
Last edited:
But its technically allowed to claim those numbers, so, there you have it.

Are they? Not sure this has been challenged legally in any way. They are getting away with it because they can at this point. If someone were to object to it as false advertising, a court may decide they are not technically allowed to claim that.
 
In the PS5 equation, power is fixed. So there are 2 variables to contributing to that power draw. 1) The amount of active compute and 2) frequency. So as you increase active compute, you decrease frequency and power remains constant. As you decrease active compute, you can increase frequency. This fact has nothing to do with the cooling solution or heat since the maximum parameters defined for those are fixed.

This is true, because there was a notion that it doesnt matter how hot the environment is, boost clocks will be the same across the board. So theoretically i could have a PS5 in my cold country operating exactly the same as someone in florida on a real hot day. A scorching hot day.
 
In the Mark Cerny presentation he said "constant power" approach can push GPU clock to 2.23GHz while "constant frequency" approach will
have trouble to maintain 2GHz.

Is there any reason why it happens? If the cooling solution is capable of 2.23GHz under heavy workload, why will there be problem to use
"constant 2GHz" even there is some power spike?

hmm the reason is that there is a power limit that the silicon can handle safely before it runs into a thermal limit problem. Whether that is through the wall or a battery source (laptops). I also think there is a budget of power between CPU and GPU and that is with or without saturation for instance.

I think traditionally on APUs, the priority for power was towards the GPU. So if the GPU started ramping up really hard, the CPU would start to get starved of power. In this scenario there is a 2 way shift, so if the CPU needed that power, it could borrow from the GPU.

Both CPU and GPUs have their own specific power curves, it should look something like this taken from AMD smartshift. You'll see wattage on the X axis and GPU performance on the Y axis.
There is sweet spot of performance that hugs the upper left corner, this is where you are getting the most performance with the least amount of wattage. As you want more performance you incur diminishing returns to get that.
As you can see on the image, each additional 25W of power the GPU gets in power, it gains a little less performance than the previous 25W shift.
csm_AMD_Ryzen_Mobile_Tech_Day_Breakout_Session_Performance_Optimization_03_feb15ab184.png


So with smart shift, you're basically setting a power bound and shifting the power around for the task. I see smartshift as a way to gain performance, but I don't think it necessarily evades the issue if the CPU and GPU are in heated contention for power.

If look at this Radeon Tool that allows you to set your frequency/wattage for your RDNA card:
DH020-05a_0.png


You'll easily see on this graph that PS5 sits just outside the axis of this graph at max cap.
But if you look at how the graph is moving exponentially upwards it's eating a lot more power once you get past the middle dot to continue to obtain more frequency. Which I think is around the 1350Mhz mark on this graph. How ironic it says 1825 for the limit set here. If we were to follow an imaginary line to 2200, the power would likely be over 1400 mv. That's nearly 2x the baseline power level.

The PS5 graph will look something like this, there will be a sweet spot somewhere in there that giving back some clock rate will be a super boon for the CPU to do it's work.
And somehow, if you're running the GPU hard at max cap rate, it must be pulling a lot of resources away from the CPU. The graph will be different of course with RDNA 2, but I have a hard time believing it's going to be that different.

Otherwise you've just got this super beefy PSU pumping out all sorts of huge power to both CPU and GPU.
 
Last edited:
hmm. the reason is that there is a power limit that the PSU is rated to provide up to before it can no longer provide any more power.
Whether that is through the wall or a battery source (laptops).

I'd venture to guess the power quota is actually set based around an overall thermal limit and not anything to do with how much actual power can be applied.
 
I'd venture to guess the power quota is actually set based around an overall thermal limit and not anything to do with how much actual power can be applied.
Yea I also assume it's going to be around the TDP. If the CPU and GPU are pulling hard eventually you'll hit a thermal limit and you need to start to downclock somewhere to save the silicon. Which is why both can't hit their max cap frequencies unless you're got this really slick cooler I assume. I'll make an edit
 
The problem is, unlike on a PC, where boost frequencies work based on temperature and so are helped by better cooling solutions, the PS5 has its clock targets fixed based on power draw. So even if you were to liquid cool it or something it would still downclock itself the same as it would using the stock fan. The only difference would be the actual temp the chip is running at.
 
I've said this before on here, but there continues to be a lot of confusion regarding the dynamic clock situation. I think Sony made a good choice by implementing it, as it smartly improves performance in all scenarios within their power and heat envelopes. However, while I have no issue with stating its maximum clocks, advertising it as a 10.28TFLOP GPU is pretty disingenuous since it's technically incapable of such performance.

I've heard mention here that maybe it can do it, but potentially only for short periods or something. That's also not true.

The PS5 leverages a dynamic clock to reduce power consumption as more of its overall compute resources are leveraged. Or as Cerny put's it, the more "demanding" the game is. It's only capable of max clocks when at some unknown percentage of ALU usage, not 100%. So it's incorrect to multiply it's entire ALU capacity at max clocks since that isn't something it can or would do. In order to determine the actual max TFLOPs number, we'd need to know the clock speed it's capable of sustaining with ALL its compute active.
PC GPUs have been advertising with the maximum theoratical TFLOPs at the max boost clock, so as CPUs. The metric does factually convey the attainable performance of that particular IP (albeit being conditional), and they did state that variable frequency applies — at least in their official blog — with FLOPs clearly being a derived metric of the "up to 2.23 GHz" figure. So from a wording standpoint, I don't see any disingenuous of Sony about it.

In order to determine the actual max TFLOPs number, we'd need to know the clock speed it's capable of sustaining with ALL its compute active.
By the nature of chip-wide DVFS power management, there is no "actual max TFLOPs number", because the whole point is that you deliver power to the place that demands it the most. The most one can argue is that the chip is more unlikely to sustain max clock in both CPU and GPU when fully loaded, and that's about it..
 
Yep. You got it now. In my opinion, it's mostly about marketing, and making the numbers "look" closer. However, there are real performance benefits from what they are doing and hurts nothing so why not. Credit must be given for that. But it's still sneaky at best.
 
PC GPUs have been advertising with the maximum theoratical TFLOPs at the max boost clock, so as CPUs. The metric does factually convey the attainable performance of that particular IP (albeit being conditional), and they did state that variable frequency applies — at least in their official blog — with FLOPs clearly being a derived metric of the "up to 2.23 GHz" figure. So from a wording standpoint, I don't see any disingenuous of Sony about it.

The difference when comparing their calculation to the PC space is that PC max boost clocks are technically sustainable with proper cooling and environmental factors since the clock isn't deterministic. The PS5 is going to knowingly downclock under a given power load and hence would never be capable under any circumstance delivering the quoted compute. It would be fine to say up to 10.28 TFLOPs if there was a scenario in which that was actually possible.
 
The problem is, unlike on a PC, where boost frequencies work based on temperature and so are helped by better cooling solutions, the PS5 has its clock targets fixed based on power draw. So even if you were to liquid cool it or something it would still downclock itself the same as it would using the stock fan. The only difference would be the actual temp the chip is running at.
i guess those liquid cooler mods won't be a thing for PS5. Interesting to note.
 
Back
Top