-
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
That, or a PlayStation Meeting aren't going to be possible for months and months. Can't have a bunch of people together in a space going hands on with hardware.
There's a big logistical issue with this reveal now that they've waited so long. Probably going to get some sort of low-key livestream...
Your current consoles dont have a change of frequency they run at a fixed point and they seem to have managed fine its more the norm and if as we expect the new gen is similar to now it will be only be 7 years.
User banned (1 day): Trolling. Platform wars.
That's my point he's claiming, theorizing. Just say we have a system that can do x and be done with it. Using peaks that aren't possible to be operating simultaneously is just causing confusion. Their plan is to confuse until release, giving hope to fans there is some secret sauce to make up a performance gap.
sorry for being so dense, but doesn't this mean that they saved some information form their talking with cerny just to make a second video after their previous kinda vague analysis??
(again sorry if i understood this wrong)
Yes. Smart shift is different than variable frequency. Smartshift gives some power from CPU (without downlocking it) to GPU when CPU is not fully used.
"has not" as in Sony hasn't said anything about it, not that the PS5 doesn't have it.
This is not 20 GB/s of bandwidth from the GPU or the Onion bus linking CPU and GPU. This is 20 GB/s from the unified memory.
There's details in the article. I found it odd they didn't talk about it in the video though.
No but the number one reason is most likely cost and maybe BC.
I was just asking. I thought Cerny presented variable clocks as an advantage like this. Then why did he choose variable clocks if fixed clocks don't represent any problem for the hardware then? I mean the CPU clock speed is already low enough so i don't see why would have to downclock it to make more room for GPU clock. At least CPU clock should have been locked imo.
Agreed. My 9900k can run at balanced (downclock at idle) or full power modes. Despite the processor being clocked at the highest level on one of them the load on it being minimal meaning it is not actually running full tilt the whole time. Long story short; Full clocks are not full load.
He explained why in the GDC presentation. Fixing clocks = leaving performance on the table in many workloads, because of the need to be conservative with clocks in case of the 'worst case' workloads that would melt the system otherwise.
I think they're pushing the GPU clocks rather high and at the top end and this is a good way to get there when needed.
...but really, running full clocks doesn't mean full load. It's not the same as utilization. So it'll have no impact on longevity.
So am I missing something or why did the video not mention basically the SSD at all? I mean if there is one thing that sticks out it's the SSD in the PS5 and it was probably what Cerny spend most of his time on during the presentation.
God how many pages had people arguing about this? Even though Cerny said in the video itself that he believes it will run at those clocks most of the time.
My interpretation, and others can correct me if I'm wrong, is that when iddle the GPU and CPU will clock down when possible, so if a current workload asks for both to run at max clockspeeds, the system will use any downtime to save power consumption and thermals. If the system can't find any workload or time it can reduce clocks, it will use AMD Smartshift to make sure GPU can keep with the needed clocks.
I guess that's when the comment "most of the time" comes from. If at some point those measures aren't enough the chip will throttle to reduce power consumption enough time.
I thought the testing done on different CU/Clock configurations at the same compute level was interesting.
Since Mark Cerny's PS5 talk there's be much discussion on Era about how "narrow but fast" was potentially more performant that "wide and slow." His example being 36CU at 1GHz may be faster than 48 CU at 750MHz, with both being 4.6TF
DF provided a very limited glimpse into how these configurations may play out in practice (12min 53sec). It's only one game, and one configuration, and I expect the results will vary on a game by game, engine by engine basis. The gist though was that in Hitman 2 the 5700 XT (40CU at 1890MHz, 9.67TF) was about 2.5% faster than a 5700 (36CU at 2100MHz, 9.67MHz) - with all other components in the system being the same. It will be interesting to see how results play out on more games.
Since Mark Cerny's PS5 talk there's be much discussion on Era about how "narrow but fast" was potentially more performant that "wide and slow." His example being 36CU at 1GHz may be faster than 48 CU at 750MHz, with both being 4.6TF
DF provided a very limited glimpse into how these configurations may play out in practice (12min 53sec). It's only one game, and one configuration, and I expect the results will vary on a game by game, engine by engine basis. The gist though was that in Hitman 2 the 5700 XT (40CU at 1890MHz, 9.67TF) was about 2.5% faster than a 5700 (36CU at 2100MHz, 9.67MHz) - with all other components in the system being the same. It will be interesting to see how results play out on more games.
Napkin maths: Assuming you're using the compression and can sustain 9Gb/sec, and that decompression is instant (not true, even for dedicated hardware), you're looking at 90Mb for a 10ms burst. The PS5's SSD is fast, but not that fast.
Is it really trolling though? It seems everyone except sony is trying to explain the ps5 and it's capabilities.
I didn't watch the video, but a good chunk of the text article talks about the SSD.
Yup. Everyone choosing to forget the sentence right after the rest is amusing.
Incremental performance improvements?
Sure if you're talking about VRS Tier 1. But VRS Tier 2 - only currently supported by Nvidia's Turing cards, and confirmed for XSX - is a whole different ball game.
It makes a significant difference to performance in addition to the other features you mentioned. And has less visual degradation impact. I've no doubt PS5 will support its own version of this as it's likely a standard part of RDNA 2.0 architecture.
Yes DF, the entire crew, are awesome. everyone join their Patreon!!
This is what Cerny meant when he said they were testing top 100 PS4 games with the boosted clock.
Don't believe it's anything to do with BC.
PS4 Pro has more CUs than PS4 it didn't have problems with BC.
technical question, does the 3d audio chip steal some power\computage performance in order to deliver that sweet quality or it's like the ray tracing chip on sex that gives the machine 13 more tf for rt stuff but doesn't steal anything from the 12 tf?
sorry if it sound stupid, but from a guy who absolutely doesn't care about audio i want to know if this stuff actually steal resources dedicated to graphics\framerate during gaming.
sorry if it sound stupid, but from a guy who absolutely doesn't care about audio i want to know if this stuff actually steal resources dedicated to graphics\framerate during gaming.
Seems like a megaboost. What's the source of this image. Just want to read about it.Incremental performance improvements?
Sure if you're talking about VRS Tier 1. But VRS Tier 2 - only currently supported by Nvidia's Turing cards, and confirmed for XSX - is a whole different ball game.
It makes a significant difference to performance in addition to the other features you mentioned. And has less visual degradation impact. I've no doubt PS5 will support its own version of this as it's likely a standard part of RDNA 2.0 architecture.
3DMark Updated with New VRS Benchmark for Tier 2 (Only Available on NVIDIA Turing GPUs)
UL Benchmarks announced to have updated 3DMark with a new Variable Rate Shading benchmark that uses Tier 2 VRS (only on NVIDIA Turing GPUs).
wccftech.com
I assume the latency is top-to-toe, including decompression, but maybe decompression is not factored in there.
But - 90MB per frame would be a lot! That's a few 4K textures worth of texels for example. Or could be enough to leave a number of classes of data to on-demand paging which had to be buffered before.
And the audio IMO.
And we....still.....don't know the form factor.
Obviously MS and Sony clearly had different design goals in mind. Someone in the speculation threads hoped this would be the case, and I am too.
It will make certain types of debates more interesting now...lol.
Just hope ppl that's extremely excited about the Series X can keep their excitement there and not in console warring.
whynotboth.gif
Thanks3DMark Updated with New VRS Benchmark for Tier 2 (Only Available on NVIDIA Turing GPUs)
UL Benchmarks announced to have updated 3DMark with a new Variable Rate Shading benchmark that uses Tier 2 VRS (only on NVIDIA Turing GPUs).
Read the article and watch the video. There's much more to it than that.
I think he's saying that the GPU and CPU can both run at the full clock the entire time - it just depends on the power budget. Some instructions are more power intensive than others, like AVX instructions on a CPU, and whatever something like Furmark does on a GPU. So, devs will have the option to do really power intensive stuff - they just have to pay it back, so to speak.
There's way too much energy spent detailing dynamic clocks, power budgets, sacrificing CPU for GPU clocks, prioritization and differing workloads. Why go through all this trouble saying you're using AMD's APU power diversion tech, and then say:
"both CPU and GPU can potentially run at their limits of 3.5GHz and 2.23GHz, it isn't the case that the developer has to choose to run one of them slower."
I think this is what is meant by potentially running the APU maxed out. The kind of game that will advertise native 4k120fps will likely have a light workload that can push both CPU and GPU to max clocks while staying under budget on power.