Yeah, that was strange for DF to ignore the ssd part. At the beginning of the video they said they're going to talk about the small details, but ignored one of the major parts of Cerny's presentation.
-
Ever wanted an RSS feed of all your favorite gaming news sites? Go check out our new Gaming Headlines feed! Read more about it here.
PlayStation 5 uncovered: the Mark Cerny tech deep dive [Digital Foundry]
- Thread starter Wollan
- Start date
Yeah, that was strange for DF to ignore the ssd part. At the beginning of the video they said they're going to talk about the small details, but ignored one of the major parts of Cerny's presentation.
I gotta say coming to this thread and realizing that some people forgot that games are not a continuous series of scripted events that have absolutely quantifiable resource requisites per frame. There is always performance overhead available to account for player agency and consequent variability, failing which we witness framerate drops and/or screen tearing.
Between things like dynamic resolution (and hopefully some tier of VRS), "smart shift" makes ample sense to divert processing power from moment to moment gameplay based on the tasks that need to be executed. After all, games on console operate at fixed framerates which offers the system driven power diversion mechanism the opportunity to check whether this decision might cause performance to drop below the target threshold, every time.
Well if it can work that way i can see a lot of 3rd party devs doing that since that how things are normally done and cut back where need be.
what do we know of how AMD does RT? Genuinely curious because normally you'd expect performance to be not directly related to CU number, but a function of CU number and frequency. If they're set up for RT differently it'd be interesting to know
Agree with your other points - there are caveats that devs may need to balance. *may* - some may not have any issues with clocks but if they do, they'll need to make a call on where to prioiritise power. Drop in actual performance shouldn't be too big, but it is one more thing to consider that they don't need to worry about on XSX. Probably not a huge issue - multiplatform devs are more likely to have flexible engines designed for multiple PC specifications so some adjustment shouldn't be a big thing
It's not.
Workloads and clock speeds are not the same thing. You can have a workload that means having to run the CPU at a lower speed in order to maintain the top GPU clock, just like you can have a workload that allows you to keep both the CPU and GPU at max.
I don't think its anything nefarious like that.
Its more along the lines that last gen DF showed most games being comparable between PS360 with a minor-to-mild edging going to 360 due to implications of shared RAM and difficulties in optimisation for cell. This gen however was a much more significant gap with PS4 widely outperforming the XB1 and it reflecting badly for MS in the DF face-offs. we all know the result of that.
MS decided to go redeem themselves with the core gaming community by chasing power again with the 1X and being very open with DF as a way to win back hearts and minds of. I doubt there is anything more nefarious going on than usual PR that targets a section of the audience MS want to appeal to.
In a benchmark like that, the benchmark can be bound by the CPU, GPU and memory. We can rule out the CPU as they usually use a powerful CPU when benchmarking. That leaves us with the GPU and/or memory. In that case, if a GPU clock frequency increase of x% doesn't result in a x% performance increase, then it's because we're memory bound or that the actual GPU clock frequency doesn't increase as much as the overclock utility implies or a combination of both. If we at the same time know that a memory clock frequency increase of y% doesn't result in a y% performance increase, then we know that it's a combination of both and if we know that a memory clock frequency increase of 9% results in a 1% performance increase, then we can deduce that the discrepancy in performance increase is mostly due to that the actual clock frequency doesn't increase as much as the overclock utility indicates. The 5700XT has a variable clock frequency and unless you know what the average clock frequency is during the benchmark, you will most likely draw the wrong conclusions. To complicate it further, the average clock frequency varies depending of the characteristics of the workloads of the benchmark, which means it varies from one benchmark to another or different resolutions of the same benchmark.
Console Top Trumps! More and higher numbers better.
Very few people commenting in here are working on games for these machines, very few understand conceptually what's going on with PS5 in particular and a number of folk are just being disingenuous.
Rich definitely ensured 100% that the core clocks were in fact stable between his tests and not dipping down below the values he listed in the video. They were locked.
You've been wrong too sometimes, it's not that important.
Then if a memory overclock of 9% only results in a 1% performance increase, where do you mean we're bound?
The heavier the workload, the greater the need to lower clocks right?
Which is what is interesting about it; the more complex the game the less clock speed(s) it will have access too.
The "Catch 22 APU."
The tests Richard discusses using two GPUs at 9.67TF -- one with less CU's and a higher clock and one with more CU's and a lower clock -- are definitely of interest to me. I hope he's able to make a video with quite a few different sample cards. 4 extra CU's at lower clocks gave a 2.5%+ boost in performance.
I know? You're preaching to the choir here...
I'm responding to a user who asked if the set power limit restricts both the CPU and GPU from running at max clocks concurrently. He's directly asking about clock speeds.
Does it restrict them? No, we already have explicit confirmation they can both run at max speeds simultaneously.
Can it? Yes, we already know that a decrease in power for a particular component, based on workload, can cause a frequency reduction.
Thank you for your explanations. Very informative. You have the patience of a saint.
Yeah. Because for me it seems like for Sony they went for a model of getting the best bang for your buck when it comes to performance. Like the 100% all the time is an absolute waste of system resources if it doesn't call for it, and I don't know any game that is like that vs games where it's almost and up and down and from what Cerny was getting at is that the devs can know exactly what performance they are getting in every scene on every PS5 and distribute workloads CPU GPU or both based on the needs on the tasks on screen.
I was exapanding on the point, not countering you.
Thanks! Just trying my best.
Somewhere else in the GPU. RDNA is good, but not perfect of course. They sell it at specific frequencies, with certain cache amounts, with a certain amount of ROPs, TMUs, with a certain memory bus, memory speed, etc. for a reason of course.
Perhaps, and you may have noticed this, Sony has been less forthcoming with everyone with details about PS5? Whereas MS has been notably more hands-on and open for discussion and in-depth technical detail. The fault for that lies not with DF/Eurogamer, but with Sony. Why not continue their partnership with Wired? Sony started the next-gen ball rolling in April 2019, mind you. Many questions, few answers.
We at least know that RT is dependent on CUs as you noted even though frequency would play some role, but my general point is that no frequency advantage here is going to overcome 16 more CUs available on the Series X with greater memory bandwidth too.
So from my understanding, the variable clocks are just a way to ensure efficient use of the CPU and GPU depending on the workload? That doesn't sound concerning and I have yet to see any devs express any sort of dread regarding the PS5 architecture. Not like it'll matter once the games are shown off. If anything, PS5 will probably be easier to develop for than the PS4
It doesn't make sense. If the GPU clock frequency increase is x% then all the units of the GPU will run x% faster and all the workloads not being memory bound should run x% faster.
MS has been more reserved privately than Sony, but more open publicly.
They are just different strategies.
Theoretically the answer is NO it cannot but neither could current gen consoles or series X, Cerny explains that if you define max load as maximum possible transistors working at once at the highest clock for the prolonged time (as unrealistic as that theoretical scenario is) it would simply shut down the console due to overheating.
gundamkyoukai
Matt
riotous
modiz
chris 1515
When I say PS3 vibes, I'm talking about compared to the PS4, which was much more "off the shelf". Here they're doing a lot of more customization on their system, which I think is cool, it just seems there's a lot of theorizing about what this certain feature can and can't do. The Cell was notoriously difficult to program for and that was a different era at Sony. I was leaning more towards the overall system configuration compared to their competition. It's similar to PS3 and 360 in that respect, in my opinion.
Matt
riotous
modiz
chris 1515
When I say PS3 vibes, I'm talking about compared to the PS4, which was much more "off the shelf". Here they're doing a lot of more customization on their system, which I think is cool, it just seems there's a lot of theorizing about what this certain feature can and can't do. The Cell was notoriously difficult to program for and that was a different era at Sony. I was leaning more towards the overall system configuration compared to their competition. It's similar to PS3 and 360 in that respect, in my opinion.
Thanks for the explanation. Appreciate what you, Dictator and DF/Eurogamer crew do.
Yeah, isn't it great?
Remember, Matt gave the nod to the Series X in an early speculation OT..with an educated guess....
So don't start doubting him now, ok?
True, I actually did not think PS3 would ship with a hard drive becouse Ken Kutaragi told me it would not.
There has not been much research done on it for RDNA, but you should spend some time to look at GCN tests made regarding core frequency and memory speed. It also did not show linear scaling at all like you mention.
Same thing here is all.
I wasn't fishing for a definitive answer like this, but I can't help noticing the wording.
You mean they will end landing on, meaning that right now they aren't? And why is that?
And again the wording here got me thinking.
What tradeoffs Ms would need to do to have the same system. And it would be possible to have the same specs and the boost system or do those trade offs would have reflected on the specs themselves?
I would think that is also because there showing off DX12U to the public .
So using there next gen system is a perfect use for it since it also good marketing .
it's not overcome, frequency just lower 1.44x advantage to 1.2x
This is hard or not hard to use by the developer. Here it is easy. I pretty sure having fast load time needs work but not difficult things.
After the other advantage, there are first party to show what the console is able to do. Sony has multiples very good studio able to show the strong point of the console. And with commercial and critical acclaim.
It's still the same fundamental basis. x86-64 architecture on the CPU and a custom GPU using the same architecture as desktop products, same memory type as desktop GPUs, etc.
Cell was a completely different, bespoke architecture. It was different at a fundamental level.
Cerny talked about this in the presentation, about being easy to learn but having room to master. They have a common base architecture with PCs and XSX, but bespoke add-ons that developers can learn to use over time to eke more out of it.
Which is actually one of the reasons I'm more excited about the PS5 than I ever was about the PS4. Developers also seem to be on board (unlike at any point prior to the PS3's launch).
BTW, I elaborated about this the other day. Not even under liquid nitrogen cooling to -50 degrees celsius does the clock frequency of the 5700XT seem to be stable. See this link, https://youtu.be/JS9in-RHbjw?t=2172.
This was a great deep dive. I too wish that Sony was more forth coming about the actual feature set of the GPU though. I'm guessing they're waiting for AMD to reveal the full specification of RDNA2? Hopefully they reveal their box and have the full teardown soon.
That's strange, but it lines up with the other reports about Sony being open with devs
I don't get it , it sounded like you just described a situation where the GPU can only run that fast if the CPU is running at a lower speed, but also a a situation where that doesn't matter and they can both run at max speed.
Rich made a video in 2019 looking at GCN vs RDNA scaling and comparing the two, that is neat.
But you can just read most of the great websites out there like Computerbase.de and take a look at any of their reviews for GCN cards like the R290x, FuryX, Vega 64, or Radeon VII and checkout their overclocking section. It will show unbalanced results when changing core and memory frequency.
There is a reason Rich stopped his tests at 2100 mhz.
But what would make a difference compared to the current 15% or so TF difference? Wouldn't RT be in line with that (until we know more about how it works)