Support NeoGAF

[Perrott]

Actually this would suggest a GPU clocked at 3.23 Ghz for PS5 BC. But still the 300 TOPs does not fit with the 33 tflops number

Would you mind to explain that in a little more detail? As in the process that led you towards that conclusion.

 

[Physiognomonics]

Would you mind to explain that in a little more detail? As in the process that led you towards that conclusion.

What I did was 2.23 ghz * 1.45 = 3.23 ghz. But still the whole thing seems off. Why would they downclock so much for PS5 Pro mode?

 

[chilichote]

So on 6nm? I don't believe that.

 

[jroc74]

Kepler chimed in on that MLiD video....interesting.

 

[Panajev2001a]

So on 6nm? I don't believe that.

Why not? 4nm and 5nm nodes are expensive and mostly monopolised by big companies the like of Apple (before they jump on 3nm).

 

[Panajev2001a]

What I did was 2.23 ghz * 1.45 = 3.23 ghz. But still the whole thing seems off. Why would they downclock so much for PS5 Pro mode?

Power consumption, they put a gazillion mode CUs and clock has suffered at that manufacturing node. It is still pretty high clocks for the GPU though, just not as high as PS5, we will see if it does pay off (it is a much newer GPU architecture so we will see if it can be more efficient per clock).

 

[Bernardougf]

Would be more cost-efficient for Sony to just release a PC version of a game.

Yeah but this is the point right ? If sony wants to release day 01 games on PC and want to remain relevant in the hardware business they must compete with at least the mid-high end PC.

This dream that people have that sony will launch day 01 games on PC and still sell 499 cheap console boxes with week hardware as usual is going to crash hard and fast.

Is like looking at this whole xbox debacle and learning fcking nothing

 

[rofif]

This generation is by far the worst. We're playing last gen games in 60-120 fps. And now we're supposed to buy a pro because what? 4k instead of 1440? You can barely see the difference if you're like most people sitting on the couch and playing on the TV. What a joke.

This is all true.

 

[chilichote]

Why not? 4nm and 5nm nodes are expensive and mostly monopolised by big companies the like of Apple (before they jump on 3nm).

But the die would be significantly larger than the regular PS5 and would require significantly more energy and better cooling. Definitely possible, that's not a question, but somehow I don't believe it.

If both consoles are produced on the same node and Sony manages to turn all the dies that are not Pro-ready due to defects into regular PS5s, the costs for the new structural process might be paid off more quickly. But I'm not an expert^^

 

[Panajev2001a]

But the die would be significantly larger than the regular PS5 and would require significantly more energy and better cooling. Definitely possible, that's not a question, but somehow I don't believe it.

If both consoles are produced on the same node and Sony manages to turn all the dies that are not Pro-ready due to defects into regular PS5s, the costs for the new structural process might be paid off more quickly. But I'm not an expert^^

OG PS5 launched on 7nm, it is just the slim in 6nm right?

Yeah, it would mean a bigger chip edging to the size of the XSX one or bigger. The biggest problem would be the GPU new features backported to 6nm if they were to be designed for 5nm chips or below. It would explain why they did not move the Zen4c or something too though.

 

[Physiognomonics]

This generation is by far the worst. We're playing last gen games in 60-120 fps. And now we're supposed to buy a pro because what? 4k instead of 1440? You can barely see the difference if you're like most people sitting on the couch and playing on the TV. What a joke.

What? People are definitely noticing those ugly FSR2 720p games upscaled to 1440p. A DLSS lite solution on a Pro model would be game changer for many.

 

[winjer]

Why not? 4nm and 5nm nodes are expensive and mostly monopolised by big companies the like of Apple (before they jump on 3nm).

But there is a big difference between N6 and N5, despite the name.
N6 is just an improved N7. While N5 is a whole new process node, with significantly less power usage and higher transistor density.
RDNA3 does use N6, but only for the smaller chips. Chips like the 7700XT and 7800XT, use N5.

What might happen is that the PS5 Pro uses chiplets. For example for cache, in N6.

 

[Killjoy-NL]

Yeah but this is the point right ? If sony wants to release day 01 games on PC and want to remain relevant in the hardware business they must compete with at least the mid-high end PC.

This dream that people have that sony will launch day 01 games on PC and still sell 499 cheap console boxes with week hardware as usual is going to crash hard and fast.

Is like looking at this whole xbox debacle and learning fcking nothing

True.

Point was more that some gaffers argued that PS4 Pro was released to prevent people moving to PC, but the situation today isn't the same and if that would be the reason, it wouldn't make sense as it would be a waste of investment for Sony.

 

[ChiefDada]

Because even in real-world performance this is looking like its going to end up inferior to a 3080/4070 which I thought was the minimum bar for a worthy mid gen upgrade. Its bandwidth starved, is quite limited in terms of a compute jump compared even to something like the ps4pro, and it would be optimistic to say its going to equal nvidia in reconstruction and rt In amds first real try with dedicated blocks.

Lol nope.

I was looking around to find what the ray tracing performance of the Pro might be like using those 2-4x RT performance increase figures.

Using the 6700 XT as a baseline, we'd end up with RT performance anywhere between a 6800 XT (2x) to a 3080 12GB (4x). This is a pure ray tracing benchmark (path tracing). Obviously, we probably won't get a fully path-traced game on consoles this generation but we can perhaps extrapolate a bit using this little graph. This would make the pure RT performance of the Pro on par with the 3080 so assuming a similar workload, the Pro should lose about as much performance as the 3080 12GB when turning RT features on in a best-case scenario. I do think the ray tracing on PlayStation is more efficient than on AMD GPUs on PC though. This might seem disappointing but do note that in a best-case scenario in a path-traced benchmark, the Pro would be a bit faster than a 7900 XTX. That's quite good I'd say.

A bit shocked at how bad AMD is at ray tracing I gotta say. Also, the 4090 is a fucking monster.

Interesting find but these numbers look so weird. When has a 4060 ever been anywhere near 2x faster than 6700xt in an RT game? And I see this was tested at 1440p. More weird.

Edit: Ah ok nvm, fully PATHTRACING. So this is probably more like that Portal RTX situation where that AMD to NVIDIA delta was insane due to significant less raster as you mentioned.

 

[Fafalada]

Are we 100% sure about this? I don't think a 60% boost in TF

But it's not a simple 60% boost - the dual-issue is far more likely to get actual utilisation on a console - as are other architecture targeted optimisations.

The whole thing is odd and numbers do not fit. 300 tops 8bit is a lot and suggest 2450 Mhz GPU clock which doesn't fit with the rest.

Assuming it's still variable clocks - that could explain some variance.
Eg. maybe running execution units for integer workloads shifts enough power-budget to get the clocks to go higher.

 

[ChiefDada]

The whole thing is odd and numbers do not fit. 300 tops 8bit is a lot and suggest 2450 Mhz GPU clock which doesn't fit with the rest. Why would they downclock the GPU when using 36CUs for PS5 BC? This should be the contrary.

For instance 2450 Mhz GPU would suggest +37 dual Tflops (+18 single tflops), not 33.

EDIT: Actually this would suggest a GPU clocked at 3.23 Ghz for PS5 BC. But still the 300 TOPs does not fit with the 33 tflops number

Perhaps protecting his source and purposely mixing in misinformation? Otherwise idk what's going on.

But it's not a simple 60% boost - the dual-issue is far more likely to get actual utilisation on a console - as are other architecture targeted optimisations.

Sure, maybe. But then that means we should actually lean more towards that 33TF figure when comparing to base PS5/RDNA2.

 

[crackajack]

But there is a big difference between N6 and N5, despite the name.
N6 is just an improved N7. While N5 is a whole new process node, with significantly less power usage and higher transistor density.
RDNA3 does use N6, but only for the smaller chips. Chips like the 7700XT and 7800XT, use N5.

What might happen is that the PS5 Pro uses chiplets. For example for cache, in N6.

imho anything else than 3nm and apple prices will make this a pretty pointless upgrade.
6 or 5nm and probably no surprising RT and FSR wonders from AMD, leaping past their current PC tech, will make this a barely noticable improvement.
With people already disappointed with PS5 and Plus Premium, Plus in general and GP and almost everything in modern gaming, everyone better should optimise their stuff, than just brute forcing some extra niceties out of current diminishing returns tech.

 

[SolidQ]

nteresting find but these numbers look so weird. When has a 4060 ever been anywhere near 2x faster than 6700xt in an RT game? And I see this was tested at 1440p. More weird.

Here is RT

Spoiler

Raster Perfomance you can calculate based on between ~7600 and 6700XT up 45%

Spoiler

 

[Embearded]

Watched the MLID video.
Does the guy really know what he is talking about?
He just divided the 67.5TF by 2, ignoring the dual issue feature, or i missed the part he mentioned it, but he clearly keeps mentioning 33.75TF for the machine.

Also, is the PSSR supposed to be a HW or SW solution?

 

[Panajev2001a]

But there is a big difference between N6 and N5, despite the name.
N6 is just an improved N7. While N5 is a whole new process node, with significantly less power usage and higher transistor density.
RDNA3 does use N6, but only for the smaller chips. Chips like the 7700XT and 7800XT, use N5.

What might happen is that the PS5 Pro uses chiplets. For example for cache, in N6.

I understand, but it does not make it more likely for Sony to use it though.

 

[Gaiff]

Lol nope.


Interesting find but these numbers look so weird. When has a 4060 ever been anywhere near 2x faster than 6700xt in an RT game? And I see this was tested at 1440p. More weird.

Almost every game is a hybrid workload of ray tracing + rasterization so raster performance still counts for a lot. This is a 100% RT workload where rasterization is almost irrelevant. That’s why I mentioned that it’s not really an indicator of performance in ray-traced games. The best you can derive from it is how much performance a card will lose when toggling ray tracing on compared to without ray tracing. The 6700 XT is actually faster than the 4060 but the moment you turn on RT, the gap starts shrinking, and the more heavy effects you add, the more the 4060 pulls ahead and if you just go HAM with the ray tracing and go path traced, then the 6700 XT just falls apart (as does the 4060 honestly). You'll never see that in non-path-traced games. Cyberpunk 2077 with PT, Minecraft PT, Quake 2 PT, or Portal RTX are such games.

Also, I mentioned it before but I think in some cases, there is some fuckery going on with AMD's RT implementation, I think something on the software side is broken.

This is Cyberpunk with just ray-traced shadows. As you can see, the PS5 still holds its own very well against NVIDIA cards but the 6700 just tanks. I think it was Nixxes or someone else who said that DRX or the RT solutions on PC weren't very flexible and like a black box when compared to PS5. Mind you, this isn't the case with every game but I don't think AMD is doing their work in some games with ray tracing. They just don't care it seems. Sony isn't that incompetent so I expect a much, much better show from the PS5 Pro assuming those numbers are real.

And there are other examples. Look at Portal RTX.

Spoiler

Do note that these are all games with heavy involvement and sponsorship from NVIDIA and seemingly none from AMD (not even sure they got drivers ready for Portal RTX). With how badly AMD is performing, you'd think the little PS5 would have no business running RT in AAA games but it does it and pretty well at that. AMD's RT is fucked up on PC. If the Pro can get RT performance on the level of a 7900 XT or 3080, that would be pretty amazing. They already get some solid results with very little hardware capabilities for RT.

 

[The ninja farmer]

imho anything else than 3nm and apple prices will make this a pretty pointless upgrade.
6 or 5nm and probably no surprising RT and FSR wonders from AMD, leaping past their current PC tech, will make this a barely noticable improvement.
With people already disappointed with PS5 and Plus Premium, Plus in general and GP and almost everything in modern gaming, everyone better should optimise their stuff, than just brute forcing some extra niceties out of current diminishing returns tech.

Can’t see Apple leaving 3nm any time soon. Not for bulk usage. And they tend to buy all or almost of the capacity on the latest node

Cost aside, I am still in the 4np camp for the new pro. I believe it’s a revised 5nm node and not an actual new node

I do believe giant chips are something that arent in Sonys current design ethos (see road to ps5 for more). Faster yes. Absolutely. Bigger. Not so much (I do buy into the 60 active compute units. Probably with the throughput doubling on certain tasks. And the shader area config will be interesting on what Sony are thinking. Much bigger than normal. Or 3 or 4 smaller. I would also think Sony and amd have developed a boost mood for previous and ps5 games that runs without a patch (like on the previous pro) and some way to map the required console cu configs to the new gpu layout (2x30, 3x20 or 4x15 compute groups)

I do think the improvements will be modest for some, but fairly huge in other areas (no ssd speed upgrade, far better rt. that sort of thing)

I was hoping for more cache at least on the cpu. But this might be one of those compatibility points that was mentioned. Which while disappointing. Does leave the area open for a massive increase on the ps6

I’m thinking more an update (and not a revolution overall) as per Sony. The best place to play ps5 games, till next gen

 

[SolidQ]

Cyberpunk 2077 with PT

That wrong numbers for Cyber

 

[Gaiff]

That wrong numbers for Cyber

What do you mean?

 

[Fafalada]

Sure, maybe. But then that means we should actually lean more towards that 33TF figure when comparing to base PS5/RDNA2.

Possibly - but this being 2024 - we all know half of people will insist there's always 0 benefit, and half will insist it's always 100%, with no ground for any discussion at all.

But being that supposed leaked docs are quoting the numbers, it should be pretty obvious they are seeing actual benefits from it, regardless of exactly how much it is.
Eg. PS4Pro had double ROPs of PS4, and that basically got ignored in the docs because in almost every benchmark you'd be so bandwidth limited it was all but useless (it helped with shadow rendering, but that was about it).

 

[SolidQ]

What do you mean?

That Fixed version

 

[Gaiff]

That Fixed version

Thanks. So it was indeed broken. Yeah, that makes much more sense.

 

[SolidQ]

So it was indeed broken.

as you can see no 15 fps

 

[saintjules]

At the end of the day with all this jargon, there wouldn't be brand new games targeting the Pro to look like new gen games, right? The point of the Pro is just to increase quality of life features that are bed ridden under the current games? As well as increased framerate and graphical fidelity to what is already there?

 

[Elios83]

Don't know if the latest leak is completely accurate, there are some inconsistencies between the 67 FP16 Teraflops and the expected raster improvement of 45%.

But I think that in general if true it is being smartly designed to address real issues and get the best possible results.
Just focus on internal 1440p resolution at 60fps, add ray tracing effects that previously weren't possible thanks to 2-3x improved ray tracing performance and upscale to 4K using AI.
This is a winning combo that would fix the weaknesses of the current PS5.

 

[Fafalada]

The point of the Pro is just to increase quality of life features that are bed ridden under the current games? As well as increased framerate and graphical fidelity to what is already there?

Yea - the same thing last gen Pro machines did.

 

[TheRedRiders]

At the end of the day with all this jargon, there wouldn't be brand new games targeting the Pro to look like new gen games, right? The point of the Pro is just to increase quality of life features that are bed ridden under the current games? As well as increased framerate and graphical fidelity to what is already there?

The Pro serves several purposes beyond just enhancing PS5 games - the new feature set targeting machine learning and ray tracing will allow some developers especially Sony first parties to build new libraries for these specific feature sets, so by the times the PS6 hits they'll already have a foundation on the newer iterations of the RT and ML feature sets.

 

[SolidQ]

only we're missed 45% raster is UNoptimized games. Like patch, unlock fps without optimization

 

[blastprocessor]

So on 6nm? I don't believe that.

Agree, doesn't make sense for 6nm, 4nm more likely.

 

[Shodai]

At the end of the day with all this jargon, there wouldn't be brand new games targeting the Pro to look like new gen games, right? The point of the Pro is just to increase quality of life features that are bed ridden under the current games? As well as increased framerate and graphical fidelity to what is already there?

If the rumors turn out correct...everything points to it being a non-intervention solution, at least as PSSR goes...which is a massive win if it simply bumps up res/frame rate. See PC/DLSS as a perfect example.

 

[Kaiserstark]

Watched the MLID video.
Does the guy really know what he is talking about?
He just divided the 67.5TF by 2, ignoring the dual issue feature, or i missed the part he mentioned it, but he clearly keeps mentioning 33.75TF for the machine.

Also, is the PSSR supposed to be a HW or SW solution?

Jeff Grubb and NX Gamer both said that they heard that the PS5 Pro will feature proprietary hardware for Machine Learning Uspcaling like DLSS.

Grubb talked about 2X RT performance with new accelerators. NXGamer also mentioned stuff like ray reconstruction.

 

[saintjules]

Yea - the same thing last gen Pro machines did.

The Pro serves several purposes beyond just enhancing PS5 games - the new feature set targeting machine learning and ray tracing will allow some developers especially Sony first parties to build new libraries for these specific feature sets, so by the times the PS6 hits they'll already have a foundation on the newer iterations of the RT and ML feature sets.

If the rumors turn out correct...everything points to it being a non-intervention solution, at least as PSSR goes...which is a massive win if it simply bumps up res/frame rate. See PC/DLSS as a perfect example.

Thank you

 

[SolidQ]

 

[foamdino]

I'm sold - this MLiD leak looks like the real deal:
* Kepler responding to it
* Not insane levels of GPU compute etc
* ML hw with actual numbers (300 TOPS)
* New type of inference PS Spectral Super Resolution (something "like DLSS" was heavily hinted at)
* Notes leaking just around GDC

I guess this will be a *huge* improvement over the base PS5 for not much cost for Sony, _and_ the interesting stuff is obviously going to be used again in the PS6

Raw numbers will be picked apart, but there's more than enough here for a really good mid-gen improvement without going mad and breaking the bank

(oh and the other thing that hints that the PS5Pro is coming, Dragon's Dogma 2 launching with an uncapped frame rate - the devs have learned from last time, if they leave things uncapped, then the mid-gen machine can boost the output with no need to patch...)

 

[Audiophile]

An out of the box thought, if it's addressable by the APU/OS (and performant enough) they could switch out the 256MB DDR3 SSD Cache for a 2-4GB DDR4 Hybrid, cordon off 256MB to remain as the Cache and then offload as much of the OS functionality to the rest it as possible, freeing up anywhere from 1.5-3.5GB of the GDDR6.

 

[jroc74]

I think people are going to be a little let down when they see the actual hardware being used and the "numbers"

I'm kinda shocked some ppl would be disappointed in 14tf-18tf....for a mid gen refresh.

 

[saintjules]

Kepler chimed in on that MLiD video....interesting.

Not only that, he didn't try to disprove anything that was discussed.

 

[VolticArchangel]

It's about time we banned this clown, Moore's Bait Is Dead, as a source, here on GAF.

The Pro better runs GTAVI at 60fps or theres no reason for me to buy It. Otherwise I'm building a new gaming PC just for GTAVI.

Game launches in September 2025.

PC + PS6 gen versions probably launch around march-june 2028.

 

[Embearded]

If PSSR doesn't require per game AI training, what does it mean for devs and even backwards compatibility?
So you think it will be usable for all games without effort / patches from devs?

 

[Gaiff]

(oh and the other thing that hints that the PS5Pro is coming, Dragon's Dogma 2 launching with an uncapped frame rate - the devs have learned from last time, if they leave things uncapped, then the mid-gen machine can boost the output with no need to patch...)

Nah, this is just dumbassery from the devs. It's some kind of vision from the future that made them do that.

 

[Audiophile]

Important not to just focus on flops. We're probably gonna see a ~60% bump on flops from compute+frontend*clocks and a ~40% bump on architecture/efficiency. Plus plenty of extra RT goodness.

Then with much higher quality upscaling we could see games with similar render resolutions but a much higher quality final image; allowing much of that extra GPU power to be spent on other aspects.

 

[Imtjnotu]

Kepler chimed in on that MLiD video....interesting.

Drop a link?

 

[shamoomoo]

Watched the MLID video.
Does the guy really know what he is talking about?
He just divided the 67.5TF by 2, ignoring the dual issue feature, or i missed the part he mentioned it, but he clearly keeps mentioning 33.75TF for the machine.

Also, is the PSSR supposed to be a HW or SW solution?

I'm not sure what you are talking. If the FP16 performance of the Pro is 67.5 TFLOPS then the FP32 rate would be half of that.


The PS5 has 10 TFLOPS of FP 32 and It's FP16 performance is double that.

 

[Anchovie123]

MLID was spot on about the RTX super pricing with many doubting the 4080 Super to come in at 999$.

If Sony actually refers to their AI reconstruction tech specifically as "PSSR" he will have earned my respect. Thats not something you can really guess.

 

[jroc74]

Drop a link?

A few ppl posted tweets in here.

I wasn't even gonna pay attention to that MLiD video until Kepler posted about it.

Some support for MLiD leak from Kepler.





Well it looks like they found a way. 2-4x RT performance uplift is excellent.

Would you mind to explain that in a little more detail? As in the process that led you towards that conclusion.