BillSpencer
Regular
It will be pretty cool for the movie-tie in game for the upcoming The Matrix reboot/sequel !
Playstation 5 [PS5] [Release November 12 2020]
- Thread starter BRiT
- Start date
The One X and PS4/Pro numbers are interesting to me. One X goes from being ~50% faster in arial to 10 slower in royal. Also, on the PC side, interesting how much 2080 loses ground in the switch from arial to royal as well. If you look at 1070 compared to PS5, it's about at the same position in both tests, about 2.5x slower than PS5, but 2080 and 1060 drop a lot of performance. There's also a wide gap between ps4 and pro between the two tests. Maybe the complexity of royal allows them to leverage RPM while arial is too simple to benefit from it. But Turing supports RPM also, right? I know Volta did so I would assume the 2080 would benefit from the same RPM optimizations.
Turing support RPM.
Comments from GN on Comments on GN Test from their followup Video @ 23:50
"That actually reminds me, as well, someone was saying 'as an engineer the material on top of the memory looks to be a cured thermal compound'
It's Not.
It's a thermal pad.
but anyways that's a different story."
It's Not.
It's a thermal pad.
but anyways that's a different story."
For a while I thought HBM would also maybe be doable for a mid-gen refresh, but having had some time to think on it more (and reading other posts here in the 10th-gen speculation thread...I gotta post some more stuff there), HBM might be too much of a shift for the coding models devs have gotten used to on GDDR memories. Plus I'm pretty sure the compression for data bandwidth on the new GPUs is much better than it was on PS4 and XBO, so they should be able to even more efficiently make use of their GDDR6 bandwidth.
HBM would also require redesigning a lot of the APU to account for an interposer, redesign the whole motherboard layout and configuration plus a PS5 Pro would probably still need liquid metal cooling and I'm not exactly sure how that'd play with integrating on an interposer for an HBM2 or HBM2E (or maybe HBMNext, which is basically HBM2E but Micron's version) memory.
It's not really about the size, but the fact that smaller nodes will cost more money. Microsoft basically gave indication of this with their Hot Chips presentation but I think a lot of that is applicable with Sony as well. So even shrinking down the current PS5 GPU to 3nm, would be very expensive (plus they may have to compete with Apple on fabs, that could artificially increase the pricing), and then we're talking about doubling the GPU hardware on top of that? That's a big ask for BOM costs.
14gbps GDDR6 probably will drop down in price, but unless 4 GB capacities come about, a 72 CU PS5 Pro would be bandwidth and capacity-starved. They could clamshell 16 2GB modules I guess, but that sounds like it'd wipe out the cost savings on the lower costs of 14Gbps chips. It doubles their capacity, though, but that's partly what would add to the pricing increase.
Personally think a PS5 Pro in 2025 is a kinda too little, too late. Sony'll probably want a PS6 by 2027 at latest, however it's possible they could go for 2026 as well. That's probably the earliest for a PS6. PS5 Pro, if they even do it, probably would be around 2023 is my guess. I doubt it'd be 3nm tho.
I kinda think so. People already say the Series X's chip is way larger than PS5's, and that's just a CU difference of 14. 52 to 72 jumps that difference up to 20. You could maybe look at it as a percentage, and percentage-wise it'd be smaller from Series X to PS5 Pro than it is from Series X to PS5. But workload scaling doesn't care for percentages in that regard, just what CUs are there, how many and therefore how many need be targeted for optimal saturation of the hardware.
Considering it as a chiplet of 2x 36 CU chiplets with one to go in a base model to cover costs of the Pro one is interesting, though. I guess they could try for something like that, it's basically what they did with the PS4 and PS4 Pro (not in terms of any repurposed chiplet designs but just the role the base unit served to offset costs on the Pro model).
But rolling generations is more something Microsoft seems to be pushing for, not Sony. I've not seen anyone from Sony talk about their future console designs in a generation-less like context, but we've seen guys like Phil Spencer refer to implicating such for Series X and S directly. How true or untrue it is comes down to perception and we'll see if Microsoft commits to that with gradual regular hardware refreshes every 2 or so years under the Series moniker.
True but only if Sony commits to concept of rolling generations, but that doesn't seem to be their mondus operandi at present. I don't think it's Mark Cerny's, either, and as long as he's the lead designer on PS consoles, I expect Sony to stick with hard breaks (aside from BC which would roll over to the next system) with gen to gen.
Part of what helps selling that is providing clear upgrades in performance from the old to the new. Personally I think PS6 (and Microsoft's 10th-gen whatever it's called) will be a lot more performant than 25 or so TF; not so much in terms of raw numbers but in technologies they'll have for offloading a lot of tasks that'd otherwise require raw compute, to dedicated silicon integrated into the designs. Although, that could be a bit less so in Microsoft's case unless they could push some of that type of work to PC space en masse (or, if they can get develop equivalent algorithms for that stuff that scale with general compute on the GPU, they could do both).
Oh goodness, I hope it's not 10 years xD. I do agree Moore's Law is slowing down, but there are some innovative approaches that can be taken to combat it. In some ways, those approaches are already being done, though some are in growing pains. Game dev time and costs are something that REALLY need to be dealt with sooner rather than later. New funding models, big growth in AI models to assist with programming GPT-style (and asset generation), will be the two biggest ways of doing so.
The latter, though, well there's an ethical line to keep in mind because you'll still need a human element present to curate what the AI creates, specialize it etc. But devs also need to ensure they don't have entire workforces replaced with automation, like what's happened to the car industry.
Anyway that's all talk going way off onto another tangent xD.
Last edited:
Nice to see him address this; saw some people questioning this shortly after the vid went out. I'd trust GN knows what he's doing when testing this stuff xD
Pretty much what I think will happen. Tho, I'm maybe a bit more bullish, for 5nm EUVL instead of 5nm, but 5nm works just as well (plus it'd be less costly). I'm also thinking chiplets may be used as you say, but release date-wise probably 2023. It's gonna be hard enough getting a base PS5 until probably at least June 2021, they honestly don't need to worry about a Pro model then or 2022 (a Slim model on 6nm might be attractive for 2022 though).
5nm EUVL, while more expensive, it'd allow Sony to get more performance but stay within (or potentially lower than) PS5's power budget; not just the GPU, but the CPU would also get some benefits in power consumption reduction. They could get more efficient NAND with lower power consumption possibly, or maybe revamp it to 3x 4-channel 256 GB modules instead of the 6x 2-channel 128 GB modules base PS5 has.
Yep, and it's also like @manux said right afterwards, too; the performance and power consumption savings will come moreso through smart/creative design choices and packaging methods rather than simply porting existing designs to smaller nodes.
I know a lot of people are probably sleeping on them, but I'm really interested in Intel's efforts in this regard. EMIB, co-EMID, Foveos, even their stuff on thermoelectric peltier plate cooling (well that one is really niche but still interesting), etc. If they could get their internal fab processes up to better par I think they have a lot of good stuff going to wring back a lot of power from AMD and put up a strong challenge to them and Nvidia in the GPU space.
Hopefully we see more on their Xe architecture in 2021, I genuinely want to see what they can do there.
Much thanks! Yeah for a PS5 Pro it simply wouldn't do. But it could maybe do something for a PS5 Slim?
Though it not helping in reducing power consumption is troubling. If anything, wouldn't a PS5 Slim on N6 just generate more heat for cooling especially if the console footprint is smaller? Doesn't sound like it'd do what many'd think it should do.
5nm might be required for that (let alone a PS5 Pro) after all.
This is a simple fill rate test where it's not at all surprising to see the PS5 coming out in front given it has 30% more raw fill rate on paper than the 2080. Describing it as a test of "computational power" and thus concluding it "destroys everything else" is a little disingenuous.
It's likely that in the same test the PS5 would also edge out a 2080Ti, but lose to a 3080 and heavily lose to anything in the AMD 68xx series. But this test isn't at all representative of a normal gaming workload.
This is partially shader based not only pixel fillrate dependent, all the part for rendering text from bezier curve. There is the shader code inside my second post. The guy writing the slug library code is the same doing the benchmark.
http://jcgt.org/published/0006/02/02/GlyphShader.glsl
And the paper explaining what they do
http://jcgt.org/published/0006/02/02/paper.pdf
And he is a gamedev, ex Naughty Dog, Sierra gamedev and he worked at Apple too.
Last edited:
D
Deleted member 13524
Guest
It might be running shaders, but it seems to me to be mostly fillrate limited.
How else would a PS4 Pro go above a OneX?
How else would a PS4 Pro go above a OneX?
Still, the output is measured in GPixels and rendering text is unlikely to be in any way shader limited. Given that the test scales as expected with raw theoretical fill rate I'd say it's a fairly safe bet to assume that's the bottleneck as opposed to this being a test of general computational performance.
Better explanation by the dev himself and at then end the result is some text displayed on screen. This is the reason they use Gpixel but there is tons of computation involved. So basically it draws text using polynomial equations. It seems like a pretty cool way to test raw compute performance.
Slug is the industry standard for UI used in game.
EDIT:
Last edited:
BillSpencer
Regular
"Still, it shouldn't count!!!
"
It looks like they removed some bottlenecks somewhere.
Five years from now, people will wonder why RDR3 runs better on launch PS5 compared to 3080
"It looks like they removed some bottlenecks somewhere.
Five years from now, people will wonder why RDR3 runs better on launch PS5 compared to 3080
"Still, it shouldn't count!!!
It looks like they removed some bottlenecks somewhere.
Five years from now, people will wonder why RDR3 runs better on launch PS5 compared to 3080
PS5 will easily outperform a 5080 too. 6900xt at 3ghz looks low end next to the ps5 gpu.
BillSpencer
Regular
The architectures are probably too close for this, and developers are (sorry) too lazy to do it, but I remember back in the SNES/Genesis(mega drive) days that games would often get a dedicated port. Even music would be recomposed sometimes to exploit each hardwares greatest strengths.
Now it would be too expensive to do it, and I don't believe the systems are that different. PS5 will be the lead platform this generation though, so it's possible that they optimise for it, and then dial it down from there for Xbox Series X; try and be less aggressive on filtrate and micro polygons and maybe try to spread the graphics across the slower, less effective Series X CUs which are superior in pure numbers.
The end result should be identical to the eye, but 'hardcore' gamers will look at analysis.
Now it would be too expensive to do it, and I don't believe the systems are that different. PS5 will be the lead platform this generation though, so it's possible that they optimise for it, and then dial it down from there for Xbox Series X; try and be less aggressive on filtrate and micro polygons and maybe try to spread the graphics across the slower, less effective Series X CUs which are superior in pure numbers.
The end result should be identical to the eye, but 'hardcore' gamers will look at analysis.
There's nothing there that says fill rate isn't a limiting factor though. It may be computationally expensive, but if you already have more than enough computational power then the bottleneck moves elsewhere.
That said, I would expect fill rate limitations to kick in in the simpler test rather than the more complex one so perhaps the answer isn't that straight forward. This still doesn't tell us anything particularly useful about general computational performance though other than RDNA2 being more efficient than Turing in this particular algorithm at higher complexity levels. It's completely normal for architectures to be better or worse than each other in specific compute tasks depending on the nature of the task. Just look at how Ampere and RDNA2 trade blows in these OpenCL tests for example:
https://videocardz.com/newz/amd-radeon-rx-6800-xt-and-rx-6800-geekbench-opencl-benchmarks-leaked
To draw any sort of conclusion for general gaming performance from this would certainly not be accurate. I'm not suggesting you're doing that btw.
In the two case they fill the screen with text. When the compute part is less complex the 2080 is a bit faster than the PS5 when the compute part grow in complexity the PS5 GPU is faster than the 2080.
I think the fillrate is not the bottleneck here. This part is the same with arial or Royal font.
EDIT: I don't believe it helps in game.
Last edited:
I do agree with you on the fill rate aspect now. I don't believe the bottleneck is straight computational performance either though. Note how the 1070 is more than twice as fast as the 1060 despite having less than 50% more shader resources while being an identical architecture. It seems the benchmarks works in mysterious ways.
Someone was going to figur this out soon or late. There will be more of those kinda of 'tests' appearing from time to time.
Similar threads
- Locked
- Deleted member 7537
- Games Industry
- Sticky