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

Thanks, this makes sense. It was how the person interviewed added the 'you can absolutely throttle to the max. We prefer if you didn’t,' That gave me slight pause. Thanks again.

 

If caring more about console wars than system design, that's true. If caring more about engineering, PS5 is possibly doing something smart that'll become the norm in future. Compared to a PS5 with that die size and without variable clocks, overall performance on the flexiPS5 should be a bit better, and had MS done the same, performance on XBSX would likely be a little bit better.

Or not. Either way, I'd prefer it if the discussion was more about comparing the engineering choices than comparing the consoles.

 

Variable clocks (dynamic frequency scaling) according to load is the norm in pretty much all modern architectures; ARM, Intel, AMD CPUs all do this as do AMD, Nvidia and Intel GPUs. This is a fundamental part of power management and thermal control - because higher clocks = more heat.

edit: In AMD technology terms, this predates current gen consoles, first appearing using the AMD PowerTune name back in 2011.

 

I think that nobody would have said nothing against if the balancing was controlled with dynamic presets by the developers, and so predictable, and more if they would have used better words instead of mumbo jumbo, so that after 5 months someone would have been able to understand how it'll work.

 

By mentioning clock speeds at all, they had to disclose the variable clocks. It definitely could have been done better. They could have emphasised that dynamic frequency scaling is how all modern non-console devices work and that PlayStation 5 was adopting this new approach to deliver a cooler, quieter console with better performance than fixed clocks.

How hard would that have been? Rhetorical question, it was apparently hard.

 

To be fair though, those with an agenda were always going to jump on 'variable clocks' and try to build a narrative of stability over it. What Cerny actually described, even going so far to explicitly state they didn't want console performance to vary across machines, was clear enough to anyone who wanted to listen. The idea of thermal throttling as something people know and associate with variable clocks shouldn't have come into it (unless one considers Cerny a liar) even if the understanding of the adaptive power wasn't well understood.

The only real, legitimate unknown is how much the processors deviate from their peak speeds and, more importantly, how that compares to if the targets were set without variable clocking. Those wanting to compare with other consoles can look at real average peak flops versus paper specs at the highest clock speed, while those wanting to compare against systems with static clocks can look at performance improvements attained by adaptive clocking.

 

The only figures that Cerny threw out was stating that reducing power consumption by 10% took a couple of percent reduction in clockspeed. At 3.5Ghz (CPU), that is 70Mhz and at 2.3Ghz (GPU) that is 46Mhz. Is 10% the cap?

 

It's the norm in the mobile and PC space (even then more so in segments with particular form factors). The peculiarity of console development, and its biggest advantage, is the ability to reliably predict performance and optimise code around it due to completely predetermined environment that is transparent to the developer. Classically, GPU performance is not constrained by CPU performance and vice versa such that you can optimise code around their limitations independently. This is no longer the case with smartshift as both processors share a common power envelope modelled by an idealised APU. How aggressive can you get in pushing the GPU before you have to start worrying about CPU performance taking a hit? That's what Jason Ronald was trying to convey in his clumsy way. Mark Cerny even alludes to it in the Eurogamer's interview by expounding about developers tailoring code for that common power envelope to maximise performance. However, in a tour de force of PR he tried to paint this additional hassle as a benefit.

 

Dynamic presets to be determined by the developer would have only served to underline how much of a hassle this paradigm actually is. Much easier to code against fixed budgets for the GPU and the CPU.

 

adaptive clock and power rates also have other problems. This can make production a bit more complicated. Every GPU & CPU combination must reach their frequencies at the desired power target. And at those high frequencies (of the GPU) I really doubt that many chips can make it. Only if those peak clocks are only really short bursts.
Problem with power-distribution. The harder developers optimize for the GPU and use it more and more intensive, the higher the GPU power draw is. This on the other hand will reduce power of the CPU. I really doubt that the GPU will be optimal used and at the same time reach the high frequencies.
Even RDNA still really uses much more power if it really get stressed. So I expect a much lighter GPU design just to reach the frequencies (e.g. less cache etc).

 

Isn't that ultimately determined by the workload? The PS5 highest power state is capped at 2.3 Ghz so the next lower power state may have a max frequency just a few percentage below the max freq. But can a workload that easily and consistently runs up against the max power limits of the console still run up against that max limit at 2.1 Ghz or below?

 

PS5 is different than PC. Variable clocks are not tied to temperature and/or chip quality. Sony uses counters inside cpu and gpu to detect load and adjusts clock speeds accordingly.

Every PS5 will behave exactly same in every condition as far as variable clocks are concerned.

 

That is where I see the problem. You can increase load of the GPU (which doesn't say about how stressed it really is) but not every calculation will max everything out. Different calculations lead to different internal load, even though the GPU can't do anything more. Even if one game creates 100% usage of the GPU at 80W (just as an example) another game might just max the GPU out at 80% but already use 150W of power. This is what makes it much more complicated for productions. More or less every CPU & GPU combination must be tested more or less with most possible load-testing and each time it must reach the same (fixed) frequencies with a fixed power target.
I doubt that many chips will make it through that binning process.

 

You can either believe what cerny said or not. Cerny said they have implemented utilization counters to cpu and gpu and temp/chip variability doesn't affect performance. He clearly said every ps5 will behave same.

 

It’s a compromise. They couldnt achieve 3 and 2ghz, now with smartshift, they can get to 3.5/2.3ghz respectively for the cpu and gpu.

Its advantageous over the former, otherwise they wouldn’t have done so. If they could attain those higher clocks sustained at the same power/cooling budged as 3 and 2ghz, they probably would have.

The dynamic downclocking will behave the same on all units despite different temps etc.

 

For sake of discussion we should call it activity level and not load. Load tends to imply weight or work. Increasing load implies a lot of compute is being done. There are ways to light up a lot of activity on the GPU without actually really doing work. Copying stuff around tends to be a pretty bad offender and does absolutely no computation.

equally there are ways to do a lot or computation with less activity level, Cerny mentioned several times that he wanted developers to adopt this mind set over time.

imo; it’s not simple. But it is a very critical topic of CS. Becoming more urgent as silicon advancements wane year over year. It is absolutely critical that we think about how our algorithms affect power usage and ways to program to reduce it while simultaneously get work completed in similar time frames. it is currently critical in fields like mobile computing, IoT, and cloud, and now we're seeing adoption here in the gaming space as well.

 

Alright, that’s true for B2C/OEM silicons that are planned to spawn a whole lineup of SKUs.

But maybe good parametric yield for one sole-and-only SKU can be alternatively attained by accounting for such production constraints in the silicon design. I wonder if something like this has been done for generations of appliances and consoles already. Ehm, maybe PlayStation? Xbox? Uhh, iPhone SoCs with platform throttling SMCs?

 

https://www.ign.com/articles/deathloop-devs-on-ps5-features-time-loop-gameplay

DeathLoop 4k 60 fps, they use some raytracing on PS5 and dualsense features. Not bad

 

Also coming to pc.

 

First paragraph:

Astute!