PlayStation 5 Pro Unboxed, 16.7 TFLOPs GPU Compute Confirmed

Ugh…. that teraflop word being tossed around like a willing hooker again!

That said, the specs are a nice upgrade from the base, so that's good.
 
HMl2riY.gif
 
I find it kinda strange they claim that dual issue stuff doesn't really apply to games, because I saw some Bend Studio dev tweet the other day more or less saying it's integral to how PSSR works.

Gonna try to find the tweet in question

Edit: Ex dev, it seems. Don't know how credible he is.


Edit 2: Might as well add the account is fake af, ignore the tweet
 
Last edited:
I find it kinda strange they claim that dual issue stuff doesn't really apply to games, because I saw some Bend Studio dev tweet the other day more or less saying it's integral to how PSSR works.

Gonna try to find the tweet in question

then it's very weird that they list the specs without dual issue fp32, and without even mentioning it.

maybe it is only used for PSSR?



Why did I keep repeatedly seeing "30+TF" thrown around :messenger_tears_of_joy:

because it's RDNA3, and RDNA3 GPUs usually get listed with dual issue FP32 included in the TFLOPS numbers. a ~33 TFLOPS RDNA3 GPU is around as powerful in practice as a ~16.5 TFLOPS RDNA2 GPU (due to architecture improvements slightly more powerful probably).
because RDNA2 doesn't have dual issue support, the numbers seem very confusing when compared directly.

Sony opted to list the specs without dual issue in this manual it seems, even tho the GPU should support it since it is RDNA3
 
Last edited:
I find it kinda strange they claim that dual issue stuff doesn't really apply to games, because I saw some Bend Studio dev tweet the other day more or less saying it's integral to how PSSR works.

Gonna try to find the tweet in question

Edit: Ex dev, it seems. Don't know how credible he is.


Firstly, that profile is fake as shit.

Secondly, that's not how teraflop calculations work. FP16 and FP32 calculations have been around long before AI upscaling was a thing.
 
Last edited:
Probably out in February next year. But the 4070 competes quite well with the Pro so I wouldn't worry on that front.
I don't understand the FP32 Float metric. I've seen double the PS5PRO and I've also seen that you have to half it in other reports.
 
Why did I keep repeatedly seeing "30+TF" thrown around :messenger_tears_of_joy:
Because ray tracing is expensive, and requires specialized hardware. When you see those outlandish numbers it is because they are saying that the GPU is doing the equivalent of that number of the Ray tracking hardware didn't exist.
 
Because ray tracing is expensive, and requires specialized hardware. When you see those outlandish numbers it is because they are saying that the GPU is doing the equivalent of that number of the Ray tracking hardware didn't exist.

Thanks for the explanation.
 
I don't understand the FP32 Float metric. I've seen double the PS5PRO and I've also seen that you have to half it in other reports.
Don't really bother with it when it comes to AMD, it's rather useless for gaming. PS5 Pro will perform like a 16.7 TF card should. Nvidia also kind of cheats with the TF system, but it's not quite the same as AMD.

Regardless of all that a PS5 Pro would probably perform between a 7700 XT and 7800 XT when it comes to regular rasterization depending on the game. A 4070 is the closest Nvidia equivalent but would probably be slightly faster in the majority of games.
 
Well done with the usual title framing DF, apparently RDNA 3 dual issue doesn't exist and the machine isn't 33.4 TF at fp32 with it. Sure it is a more 'acceptable' figure to be put besides XSX' 12.
 
I find it kinda strange they claim that dual issue stuff doesn't really apply to games, because I saw some Bend Studio dev tweet the other day more or less saying it's integral to how PSSR works.

Gonna try to find the tweet in question

Edit: Ex dev, it seems. Don't know how credible he is.

Manchester, England
I thought Bend was in the US.

And they can't even spell Smith.
Edit: There's another one lol
 
Last edited:
I love that its smaller and judging from how he's carrying the unit, it seems lighter. Very surprising yet impressive that Sony was able to make all those modifications.
 
Last edited:
Well done with the usual title framing DF, apparently RDNA 3 dual issue doesn't exist and the machine isn't 33.4 TF at fp32 with it. Sure it is a more 'acceptable' figure to be put besides XSX' 12.
It's literally the figure listed in the manual.
 
Why did I keep repeatedly seeing "30+TF" thrown around :messenger_tears_of_joy:

Different calc. method that started with RDNA3. If you use calc. method for RDNA3/4 it would indeed be double the TF. You cannot compare these TF numbers in regards of performance.

Sony stating 16,7TF* for PS5 Pro(RDNA4) is correct performance wise vs. PS5 Standard(RDNA2), but mathematically correct would be ~33TF.

*Makes sence, since Sony stated ~45% performance increase in Raserization for Pro vs.Standard.

Check out this source:
 
Last edited:
Different calc. method that started with RDNA3. If you use calc. method for RDNA3/4 it would indeed be double the TF. You cannot compare these TF numbers in regards of performance.

Sony stating 16,7TF* for PS5 Pro(RDNA4) is correct performance wise vs. PS5 Standard(RDNA2), but mathematically correct would be ~33TF.

*Makes sence, since Sony stated ~45% performance increase in Raserization for Pro vs.Standard.

Check out this source:

It may very well be that Sony didn't include dual compute in the design of the Pro GPU as they saw no need for it. They did a similar thing with the PS4 Pro, when the GPU had select features from Polaris and Vega but not all.

But in the end it really doesn't matter, 16.7 TF gives an accurate picture of the expected performance from the GPU.
 
Last edited:
Well done with the usual title framing DF, apparently RDNA 3 dual issue doesn't exist and the machine isn't 33.4 TF at fp32 with it. Sure it is a more 'acceptable' figure to be put besides XSX' 12.

dual issue has never shown to have any tangible benefits for games.
which Sony also clearly thinks, which is why they officially list it without dual issue
 
Last edited:
Possibly they could have dropped the 32-bit instructions of VOPD? Similarly to how they use the Mendocino variant of Zen2 with the lower performance FPU to save a couple mm² of die space. They haven't really resulted in much in the PC space in terms of performance in either gaming or compute.

We know they have them for lower precision modes and list them there in reveal, e.g. 67TF FP16 and 300 TOPs (likely sparse int8) so it'd be pretty weird not to do the same with the FP32 number.
 
Top Bottom