which takes us to probably being pad limited in a 5nm process. They could be forced to less than 256 bits wide bandwidth RAM. So, probably thats another reason for going 448 GB/s instead of 510.
-
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 System Architecture Deep Dive |OT| Secret Agent Cerny
- Thread starter vestan
- Start date
- Status
Yes, the down clock scenarios occur when the total power draw of the workloads would have exceeded the power limit. Cerny specifically mentioned that AVX 256 instructions were power hungry, so avoiding unnecessary use of them will be one type of power draw optimization. The key is that a variable frequency system should be able to achieve much better efficiency since 1) down clocking will only occur when the total power draw of the workloads exceed the power limit and 2) by profiling and optimizing for power draw the developer can try to avoid those cases as much as possible.
Last edited:
Hey anexanhume DrKeo could you guys explain why it seems like Sony is stuck with 36CUs. Its like they are being held hostage by previous decisions. Is this the case or just a choice Sony is making?
7nm is an expensive node. They wanted to keep the cost down at all cost, above all taking into account their precy SSD. But in the end the gamble could be right as they are not that far in specs and have made a possible Lockhart useless if they get the 399 eur price.
They'll probably go to 6nm once it's cost-effective.
I think there's a complex truth here. 7nm at the time they were planning would have been more expensive, and the yield maturation wouldn't have been guaranteed. It made sense to keep die size down.
They may have convinced themselves 36 CUs was a natural fit because it made BC easier. Keep in mind Cerny said there was a "tension" between adding new features and maintaining BC.
I think they could have done a larger GPU and figured out a sensible WGP deactivation protocol, but convinced themselves 36CU was good enough with their RRP targets.
That's the only thing that you took from the post that 'the unheard of 2.23GHz is only on PS5'?? Well I didn't say that actually. I said it happened for top of the line, overclocker cards like KINGPIN which are only handful as of yet which is correct. On the console side it will be unique to PS5 however so my point still stands. On PC side the exotic AIO coolers might make it a common occurrence but still these cards will not magically make all RDNA2 cards 2+ GHz as it will presumably still be dependent on great cooling.
Well let me iterate that previous gen consoles were working in steppings like 100%-90%-80% fan speed and dB-noise levels are not linear as well, so if the fan was spinning at 1200rpm let's say it would add/take away ~100rpm to affect a meaningful change in thermals with huge difference in noise levels. On the frequency side it is the same, decrease the clock in steps but not to something in between. These were open to optimization on the developer side as the console manufacturer enabled that but there was no guarantee the console in question would be hitting those wide berth steps based on ambient thermals, and the noise is also dependent on ambient noise.
Now the console manufacturer guarantees thermals and thus noise level will be independent of the ambient situation and the devs can target power level they want without worrying about these and do this much precisely AND the console itself will vary (not go down a step or up a step and make huge differences) only in few MHz and few rpms as it needs/requires based on the ambient to hit the targeted power. The console is in the final control to put it in simpler terms.
This is not to go against the rest of your post which I agree with, but rather to provide some evidence that these frequencies are likely to be hit most of the time: My 5700 XT can keep pretty consistently 2.0ghz with 99% GPU usage (Forza 4 Horizon benchmark running at 4k/~80fps) while drawing ~180W. Now keep in mind that this card is 40CU (4 more than PS5) and does not have the efficiency improvements of RDNA2. Incidentally, I have the manufacturer card with notoriously bad cooling and it still is quieter than some launch PS4 I heard. We will have a clearer picture when RDNA2 cards release, but people anticipating 2.0Ghz/9.2tf to be the actual performance of the system will be positively surprised. I think 2.0 was the target for RDNA1.
A general comment I'll make is that I'm excited how the console makers are pushing the envelope here with imagining different performance profiles and targets.
They're doing desktop or better clocks and thinking creatively on cooling and clocking and power strategies. I have confidence they're doing this with transparency/ease to developers in mind.
I don't know if by data driven you mean that the SSD can alleviate a bottleneck, but from what I understand motion matching is closer to a look-up table and in this case I guess memory usage could be an issue (have no idea how data hungry animations are). But for the cloth simulation I don't think the SSD is of any help at all. It is data driven in a different way, in the cloth simulation data is generated for offline training of the simulator but is not kept (at least not as is) later on.
The animation model created for the cloth simulation after training is, literally, data. You can think of machine learning as a method of dynamic compression, you're basically turning a bunch of inputs into outputs that can be reconstructed.
A heavily compressed version of the original data, yes. We are talking about a few MBs here. In contrast, the computational cost could be a problem especially if you have multiple of these models.
I'd have to find a spot for that in my shelves, it's something I'd want to display.
I've dealt with ML models (image reconstruction models) in the tens of GBs, so yeah, no. Complex models can get pretty large, and I can't imagine an animation model to be particularly simple.
These are not for image reconstruction. Examples are provided in the presentation, they are all between 20~50mb.
I'm pretty confident the PS5 will be able to do more with less. Pair it with great developer tools and I am sure we'll see 3rd party games be largely on par.
As the presentation itself talks about, it's a trade-off between memory footprint and computation (like all compression). These particular models are this size, doesn't mean they necessarily need to be.
Party?
You know that two or more people are not allowed to be toghether. I hear there is a crisis. So I have to decline the invitation. :)
Do we know if the M.2 slot will be useable for cold storage (without speed requirements) or reserved for the system ?
I'd say the main trade-off is between memory and quality. When you reduce memory you also reduce computation. If they chose models this small it's likely because they hit a computational target before a memory target.
And when you will have this for dozens of NPC, you can add this for fluid simulation and it is more data for example. ;) Data driven realtime rendering is a new field but with fast SSD we will heard of it a lot... ;)
Probably more difficult but why not? This data too and interactive...
Physics Forests: Using Machine Learning for Real-time Fluid Simulation - ACM SIGGRAPH Blog
What makes machine learning and real-time graphics a winning combination? The team behind SIGGRAPH 2017's Physics Forests explains.
blog.siggraph.org
The guy did a live demo at SIGGRAPH 2017 and read the statement in bold... A company was created with this idea.
Last edited:
What better way to fix fan noise complaints than to not have one? ;)
Of course I think Lockhart could be a portable, so maybe I'm overly optimistic about the next gen AMD power usage.
oh we still getting sacrifices for people who had their date post event?
lol good stuff
Well you can always reuse a model on several objects. For instance, you could drop several bunnies in that scene and the memory cost will stay the same but the computational cost of simulating several bunnies deforming will always increase with the number of simulations you'll run in parallel.
Obviously this is just speculation on my part and only a game dev working on this can say for sure. But I don't think an SSD will play any major role in data-driven physics simulation. At least not given what was shown in this presentation.
But why do you want to drop the same model, you want repeat the same assets over and over, this is not the goal...
The only downside is in production, they need to launch the learning process for every new objects... And this is not cheap on time but this AAA production...
Confident based on PS4 tooling and lessons learned from the PS3 era. Sony did not sit around idle before last generation and the quality difference between PS3 and PS4 on the OS and network front were an upgrade. I say that as someone that loved the XMB.
That would be about 10-15% smaller then even original 28nm PS4 APU?
Very good sign for potential pricing!
I dont buy the 36 CUs was neccesary for BC at all, is not like the own PS4 pro deativate 16 CUs for doing it.
If that's true why do most manufacturers have the boost clock on the 5700xt at right around 2.0? I think we need to consider yields also not just thermals. You may have hit the silicon lottery. RDNA 2 is shaping up to be pretty great if Sony can maintain 2.2 on what's effectively the RDNA2 version of the 5700xt.
I don't see why you would use different cloth deformation models for each NPC...
But you are right, you could have several physical processes that are so loosely related that it makes no sense to learn them with a single model. Still in absolute terms the memory footprint as shown in the slides is so low that if you even have 10 very different and complex physical processes that you are simulating in parallel, it'll take you what 250/500mb of memory? Just does not seem that high to me.
Well before we get 10, let's at least hope we get one. Having object deformation like it is shown in the video you shared would be amazing for interaction in VR. I think it would be enough of a breakthrough to warrant a new half-life :).
I don't think so. Microsoft said absolutely everything about the specs of the XSX, but they never mentioned SmartShift. IF they said everything else, why keep that to themselves? It's very likely the XSX doesn't have it.
I don't know much about OC and used the automatic undervolting setting that AMD offers. It did wonders in lowering wattage while keeping high frequency. Yeah maybe I was lucky with my card. By I also read of liquid cooled OC of the 5700 xt that can reach 2.3ghz!
Microsoft made quite a big deal about having fixed clocks, I don't think they're using smart shift.
From what I understand they need to train for every type of clothes and simulation. Here another video from last year from Ubi soft. This not one simultation for every clothes. This would be too simple...
And they explain in this video it takes a lot of time to train for each model...
One could theoretically use one model for every type of cloth material and parametrize the input, but then you'll need more modes to get the same quality for any given output - i.e. you're increasing exponentially the model complexity.
Exactly the more RAM and streaming speed, the best is the quality maybe one day the PS5 SSD will be too slow and the PS5 will not have enough memory. I think the worst is not cloth animation but fluid simulation not sure it will be usable this gen.
This was a realtime fluid data driven simulation presented live at SIGGRAPH 2017 with object being interactive.
User banned (3 days): platform warring, history of similar
I have found this account that talking about leaks and confurmations about ps5
He claims that ps5 will use RNDA 1...
twitter.com
He claims that ps5 will use RNDA 1...
blue nugroho on Twitter
“https://t.co/pIj3C2fQSh with official PS5 overclckd 2.2Ghs+=10.3TF =score is ~23K or equal to 13TF GCN RVII, but the 2.2+ is smartshift🤦 so the typical clock i bet is 1.8-2Ghz at 36CU!! < 9.7TF 5700XT score 🤦 OTOH: real RDNA2 is 30-40% way more than this score on same TF”
twitter.com
That's an Xbox fanboy account parroting old and verified false information.I have found this account that talking about leaks and confurmations about ps5
He claims that ps5 will use RNDA 1...
blue nugroho on Twitter
“https://t.co/pIj3C2fQSh with official PS5 overclckd 2.2Ghs+=10.3TF =score is ~23K or equal to 13TF GCN RVII, but the 2.2+ is smartshift🤦 so the typical clock i bet is 1.8-2Ghz at 36CU!! < 9.7TF 5700XT score 🤦 OTOH: real RDNA2 is 30-40% way more than this score on same TF”
Yep you can see that 36CU will price reduce fast .
Anex was saying they will jump to 6nm when it comes but when is the time table for that .
- Status