Current Generation Hardware Speculation with a Technical Spin [post GDC 2020] [XBSX, PS5]

[iroboto]

Ok so with Fixed clock power is always at 2.0GHz and it doesn’t work, but with their power shift it can go from 2.23GHz to (single digit percentage drop max of 9%) 2.03GHz? In this case it still means that the GPU is operating the entire time at frequency higher than 2.0GHz. Unless the GPU is allowed to swing much more and scale with workloads?


Also while they don’t want ambient temperatures to affect the performance of the chip, it still must have some type of thermal protection in case it can’t get enough airflow or something. It would probably just shut down with an error and not just 100% ignore temperatures of the chip and possibly damage it.

It likely swings well below 2.0 GHz if it requires it. This all depends on how their thermal and power is setup. The less cooling and less power, the swing will need to be greater. Better thermals and power, the swing should be less.

These swings aren't constant however, they will blip down as required and blip back up.
The question that hasn't really been answered really comes down to what the clock rate looks like when it's under sustained load, because blipping up and down is very fast, we have no real ideal of effective computational output except to benchmark.

Usually with current boost clocks, as the load goes up, boost goes down, if the GPU goes idle it's boosting back to maximum. So in some ways it's boosting through idle, which is desirable, but not an actual measure of work completed. A lot of idle time could signal to the developers that there is a lot optimization that could be done. Unfortunately the behaviour of PS5 is not really known yet. My personal speculation is that it will ride high on clocks using the traditional fixed function pipeline (3D), but with compute shader pipeline where the developers are able to saturate all the CUs in a very effective manner, the clocks will drop when they all go to work at once

A pure compute shader heavy engine like UE5, probably pushes the PS5 very hard on that front I suspect. But the results are still very good.

Yes, there will likely be thermal limits. I suppose it would shut down like any other fixed console would in the worst case scenario. But the thermals will not affect the clockspeed, that should be the workload/activity monitor.

 

[Kreten]

It likely swings well below 2.0 GHz if it requires it. This all depends on how their thermal and power is setup. The less cooling and less power, the swing will need to be greater. Better thermals and power, the swing should be less.

These swings aren't constant however, they will blip down as required and blip back up.
The question that hasn't really been answered really comes down to what the clock rate looks like when it's under sustained load, because blipping up and down is very fast, we have no real ideal of effective computational output except to benchmark.

Usually with current boost clocks, as the load goes up, boost goes down, if the GPU goes idle it's boosting back to maximum. So in some ways it's boosting through idle, which is desirable, but not an actual measure of work completed. A lot of idle time could signal to the developers that there is a lot optimization that could be done. Unfortunately the behaviour of PS5 is not really known yet. My personal speculation is that it will ride high on clocks using the traditional fixed function pipeline (3D), but with compute shader pipeline where the developers are able to saturate all the CUs in a very effective manner, the clocks will drop when they all go to work at once

A pure compute shader heavy engine like UE5, probably pushes the PS5 very hard on that front I suspect. But the results are still very good.

Yes, there will likely be thermal limits. I suppose it would shut down like any other fixed console would in the worst case scenario. But the thermals will not affect the clockspeed, that should be the workload/activity monitor.

Thanks, and this is what makes sense to me. Maybe I misunderstood what Cerny was saying, but I took it as the frequency drop would be no bigger than “single digit percentage” so things didn’t make sense.

 

[function]

Thanks, and this is what makes sense to me. Maybe I misunderstood what Cerny was saying, but I took it as the frequency drop would be no bigger than “single digit percentage” so things didn’t make sense.

Cerny was very, very careful never to give a maximum for frequency drops, though his talk indicates that it could - in extreme cases - be lower than the maximum fixed frequency levels that they were unable to achieve.

All we can be sure of is that the situation will never be worse than it would be for PS5 fixed clocks.

Personally though, and looking at the way power draw varies on PC with different loads, I expect the maximum drops on one area (CPU or GPU) to be somewhat greater than "2%" once games are being optimised to squeeze every operation they can out of these processors.

 

[see colon]

I think a year or two in and games will keep the GPU clocks near the maximum and the CPU will be the part most often down clocked anyway. If you are aiming for 4k60 with ray tracing etc, the GPU is going to limit system performance before the CPU. Actually, it might be that way from the start. If you are running game engines that also ran on PS4, your CPU is going to be able to handle that without issue, and if you are moving from pro's 1440p-ish pixel counts to real 4k, doubling frame rates, or cranking up graphical setting you will need that GPU power.

 

[disco_]

Cerny was very, very careful never to give a maximum for frequency drops, though his talk indicates that it could - in extreme cases - be lower than the maximum fixed frequency levels that they were unable to achieve.

All we can be sure of is that the situation will never be worse than it would be for PS5 fixed clocks.

Personally though, and looking at the way power draw varies on PC with different loads, I expect the maximum drops on one area (CPU or GPU) to be somewhat greater than "2%" once games are being optimised to squeeze every operation they can out of these processors.

So a ~10% clock drop resulting in ~30-40% power draw drop? As a dev, what are you doing that you're exceeding power limit by 30%?
All the recent news is making this box look like a hack job.

 

[Frenetic Pony]

So a ~10% clock drop resulting in ~30-40% power draw drop? As a dev, what are you doing that you're exceeding power limit by 30%?
All the recent news is making this box look like a hack job.

That's uh, just how modern transistors work. All of them. Hooray exponential f-max curves!

 

[iroboto]

So a ~10% clock drop resulting in ~30-40% power draw drop? As a dev, what are you doing that you're exceeding power limit by 30%?
All the recent news is making this box look like a hack job.

Power is cubic to frequency. That’s why we have hard limits on how fast we go and we started going multi-core.

So if we assume at regular frequency it’s 100W. A 10% drop in frequency. Is now 90%
(0.9)^3*100 = 72.9 W.

Nearly 30W reduction or 30% in this example with 100W

 

[itsmydamnation]

So a ~10% clock drop resulting in ~30-40% power draw drop? As a dev, what are you doing that you're exceeding power limit by 30%?
All the recent news is making this box look like a hack job.

i find your post ironic.......

maybe not ironic exactly, but something......

 

[Deleted member 11852]

Ok so with Fixed clock power is always at 2.0GHz and it doesn’t work, but with their power shift it can go from 2.23GHz to (single digit percentage drop max of 9%) 2.03GHz? In this case it still means that the GPU is operating the entire time at frequency higher than 2.0GHz. Unless the GPU is allowed to swing much more and scale with workloads?

The goal is cooler, quieter operation and pumping in "a generous" (-Mark Cerny) amount of power and letting clocks rise until the power draw meets the capability of PS5's cooling. Any given clockspeed in a CPU or GPU is not a useful metric for determining power draw, PS5's CPU could be running at 1.99Ghz and drawing more power using 256-bit instructions that the CPU running at 2.3Ghz not using those instructions and the same is true in the GPU. Not all CPU and GPU operations are equal in terms of hardware usage and power draw.

Variable clock rates are the norm everywhere and have been a while. Laptops, desktops and mobile devices have been using variable clocks to manage heat and power draw for almost a decade. Sony are adopting this approach a little latter than other devices. Mark Cerny said they're using AMD's own SmartShift technology, where PS5's implementation differs is some form of CPU/GPU workload monitor to any given code performs deterministically identically on any two PS5 consoles no matter their environment, e.g. ambient room temp.

If you look at CPU/GPU utilisation in any PC game, it's staggeringly rare for both CPU and GPU to be running at max any time, there are too many architectural bottlenecks in any complex system to every allow both processing units to reach their theoretical maximum performance simultaneously and on it's this basis the technology is based. AMD claim enabling SmartShift results in a 14% increase on GPU performance.

So a ~10% clock drop resulting in ~30-40% power draw drop? As a dev, what are you doing that you're exceeding power limit by 30%?

Mark Cerny said a couple of percent drop in clock speed results in a 10% drop in power draw. It's power draw that they want to control. Dropping 2% of the CPU's max clock of 3.5Ghz is down 70Mhz. Dropping 2% of the GPU's max of 2.23Ghz is 46Mhz. It's not a linear correlation between clock and power draw either. This is likely why Mark Cerny said "I'd expect any down clocking to be pretty minor all things considered".

 

[ultragpu]

Was this posted? Seems Geometry engine related patent from Mark Cerny.
http://images2.freshpatents.com/pdf/US20180047129A1.pdf

 

[Allandor]

If you look at CPU/GPU utilisation in any PC game, it's staggeringly rare for both CPU and GPU to be running at max any time, there are too many architectural bottlenecks in any complex system to every allow both processing units to reach their theoretical maximum performance simultaneously and on it's this basis the technology is based. AMD claim enabling SmartShift results in a 14% increase on GPU performance.

Yes, but that is only true if you have a fixed power target, that is lower than both components can reach at their max settings.
(edit: this only applies to the 14%+ through smartshift)

Variable frequencies are not really a problem like we have them on pc. But if you have fixed hardware (well thx to the variable clock speeds it is not really fixed) you may want to optimize to never go above that fixed performance target. As a developer, you know how many e.g. compute power you have as budget. Yes you must still be below that because you can't use it 100% efficient. But when designing a game, you have that specific target and you try to never go above that target, else drops in framerate will occur.
Yes you can do stuff like dynamic resolution etc, to minimize those problems, but those tricks introduce their own problems.

It is much harder to optimize for "dynamic" frequencies. You will never set your target higher, than what the system can sustain the whole time, so you will set your target a good margin below the max performance target.

On PCs this concept is not really a problem, just because most of the time you just waste potential (e.g. activating vsync and everything above 60 is just wasted) or you have lower framerates that have their own problems we (at least here in a technical forum) don't want to see.

Funny thing about that, as soon as most TVs support VRR/Freesync, stable framerates are not longer that important like they are now. Yes you still have the problem with uneven animations etc (VRR/Freesync can't do something against that, only output the frames more evenly), that variable frequencies might be a good option. But for now, we are far away that all TVs support that feature.

 

[Deleted member 11852]

Yes, but that is only true if you have a fixed power target, that is lower than both components can reach at their max settings.

Not just fixed power target, we've seen enough GDC analysis with devs showing CPU and GPU utilisation and it's pretty much impossible to get 100% out of the CPU and GPU at the same time because the workloads are never that well balanced -not unless it's a really, really simple game.

Achieving peak utilisation of one device is hard, but all cores across two processing architectures working in concert simultaneously?

If you can show me an example where that has ever happen, I'll be having hat for dinner!

 

[Allandor]

Not just fixed power target, we've seen enough GDC analysis with devs showing CPU and GPU utilisation and it's pretty much impossible to get 100% out of the CPU and GPU at the same time because the workloads are never that well balanced -not unless it's a really, really simple game.

Achieving peak utilisation of one device is hard, but all cores across two processing architectures working in concert simultaneously?

If you can show me an example where that has ever happen, I'll be having hat for dinner!

Sorry my quote was not specific enough. I meant the 14% thx to Smartshift is only valid for a fixed power target
All that only matters, if you got a fixed power target. With a flexible powertarget (like CPUs and GPUs normally have) you always get those 14% (and even more) because the CPU & GPU can always reach their max frequencies (if the cooling solution is good enough).

Variable frequencies make it just much harder optimize for a system.
And I don't mean variable frequencies on PC that occur if the CPU or GPU has not enough to do. Than they will normally reduce their frequencies until they get something to do. This could be still be the case on xbox just to save some power if the game is just not stressing everything out of the components. But if it does, you always have those frequencies you can count on.

 

[Deleted member 11852]

Sorry my quote was not specific enough. I meant the 14% thx to Smartshift is only valid for a fixed power target All that only matters, if you got a fixed power target. With a flexible powertarget (like CPUs and GPUs normally have) you always get those 14% (and even more) because the CPU & GPU can always reach their max frequencies (if the cooling solution is good enough).

Yes, this is the case. In a PC the two power inputs which would be on separate rails but in PS5 a single input. I'm not seeing how this would make a difference though, at crunch the power available is finite from the PC's PSU over two rails, and power will be finite from PS5's PSU.

Your PC is generally not focussed on power draw as much as clock speed - there being a basic assumption that the PSU is sufficient to supply power in the most demanding circumstances and can drive the CPU and GPU at their highest rated clock speeds - subject to the cooling system not imposing thermal limits. PS5 should not have thermal limits clocking down the CPU or GPU though because the cooling system will be rated for the PSU's worse case scenario.

It will be interesting to see how this all works in practise, I'll be watching GDC videos in case anybody covers it but I imagine it'll largely be a non-issue. We've only heard positive sentiments that PS5 is a joy to work with and you would think that if this new design paradigm was problematic, somebody - at least one person - would have spoken up.

edit: grammar

 

[Dictator]

Was this posted? Seems Geometry engine related patent from Mark Cerny.
http://images2.freshpatents.com/pdf/US20180047129A1.pdf

Reading through this, it seems like work related to Simultanous Multiple Projection that was introduced in NV's Pascal (and a bit in Maxwell I guess before that): you could use that to partition up screen space for lower shading.

 

[Jay]

I think a year or two in and games will keep the GPU clocks near the maximum and the CPU will be the part most often down clocked anyway. If you are aiming for 4k60 with ray tracing etc, the GPU is going to limit system performance before the CPU.

Where as I think the opposite.
Given techniques like dynamic resolution, easier to scale the load on the gpu, than on the CPU.
I'd rather a drop in resolution before a stutter in animation or physics etc.

 

[Quadbitnomial]

Reading through this, it seems like work related to Simultanous Multiple Projection that was introduced in NV's Pascal (and a bit in Maxwell I guess before that): you could use that to partition up screen space for lower shading.

It is not, SMP is taking a viewport and reprojection it several times, slightly altering the output by shifting the result along the X axis, this has now been supplanted by Multi-View Rendering in Turing which expands on it with the ability to shift each view along any axis and is position independent. Its a cheap way to achieve stereoscopy. If anything, it has more in common with MRS (multi resolution shading) from Maxell which is something PS4 Pro already supports in hardware for PSVR. PS4 Pro support has an upper limit of 5x5 rectangular regions aligned to 32x32 pixel supertile boundaries. Reading the patent it seems to cover a multitude of topics beyond just multi resolution shading, I'm guessing that is where the "under varying rendering and rasterization parameter" comes in.

 

[function]

So a ~10% clock drop resulting in ~30-40% power draw drop? As a dev, what are you doing that you're exceeding power limit by 30%?
All the recent news is making this box look like a hack job.

Well I can't speak as a dev, but on PC you can see from on-chip monitors that power demands vary greatly across tasks for both GPUs and CPUs. My CPU can be running "flat out" in a game at max clocks and pulling something like 70~80W. Or I can render something in Blender using ray tracing (my GPU is very old, the CPU is basically as fast!) and whom it's off over 110W, hitting the power limit I set and throttling down by a couple of hundred mHz. And that's with it undervolted and with hyperthreading disabled (because, well, it's an Intel).

Getting over my comically low "TDP" by more than 30% is comically easy, even without a power virus! And Sony can't get away with fudging things or have mobo makers turn off power limits on overclocking boards by default.

You shouldn't see PS5 as a hack job, because what they're doing is a perfectly legitimate engineering solution to a problem you have to face one way or another. If Sony were, indeed, limited to 36 CUs by backwards compatibility, then going faster is the only way to hit a 10 TF performance goal. And if you keep going faster, the variance of power required over time is going to become a bigger and bigger problem and you're going to go past a point where you have to drop frequency (and voltage) to keep things below your cooling and power levels. It's the same challenge PCs have been facing for years!

Not just fixed power target, we've seen enough GDC analysis with devs showing CPU and GPU utilisation and it's pretty much impossible to get 100% out of the CPU and GPU at the same time because the workloads are never that well balanced -not unless it's a really, really simple game.

Achieving peak utilisation of one device is hard, but all cores across two processing architectures working in concert simultaneously?

If you can show me an example where that has ever happen, I'll be having hat for dinner!

Indeed! And game workloads vary over frame time too, so even if you could get 100% out of all cores and all compute units at once while actually doing useful work (impossible!) would you be doing it for 100% of the time?

Even if your GPU throttled / un-boosted by 10%, if that's for 20% of your frame time, you're still averaging 98%!

 

[Kreten]

Reading through this, it seems like work related to Simultanous Multiple Projection that was introduced in NV's Pascal (and a bit in Maxwell I guess before that): you could use that to partition up screen space for lower shading.

This makes me think it’s VR related

 

[milk]

One thing I like about ps5 going variable clock as standard is that future "boost mode" scenarios will probably work way more reliably. It might be a great thing for foward compatibility.