OP
OP
Eurogamer Spec Analysis: PlayStation 5 Pro - the most powerful console yet |UP| 4/1 How Did Sony Beat Microsoft In Machine Learning For Consoles?
- Thread starter Gamernyc78
- Start date
Very surprising that a gaming tech enthusiast YouTube channel who seemed so excited for the likes of Xbox One X and Series S/X now seems so muted on a PS5 pro with RT capabilities you’ve only seen in Nvidia GPUs and ML upscaling you’ve only seen in Nvidia GPUs.
So odd wonder what that’s about
So odd wonder what that’s about
Thank you for the explanation Mr. Greenberg, I didn't realize you were still around.
Damn, can you lie just a little more and be even more dishonest? Reconstructed 4k is not the same as outputting a 1080p feed into a 4k display - Edit: When outputting 1080p content in a 4k TV, each pixel is represented 4 times to account for the 4x resolution increase in 4k.
I'm not going to be explaining several types of image reconstruction techniques, but the important is that the output, no matter what your baseline is, will always be of a higher resolution and of a higher fidelity due to reconstruction techniques. This is not simply grabbing a 1080p signal and just having it output on a 4k tv.
And you know what's even better and funnier? Reconstructed 4k can look better and more detailed than compressed raw 4k (meaning you grab a raw file and compress it to save space). Fantastic, right?
Last edited:
I completely realize this, but trying to conflate a heavily reconstructed image as 4K without notation is just as intellectually dishonest.
@Satoru wasn't conflating reconstruction with full resolution rasterization. But conflating image reconstruction with display device upscaling is intellectually dishonest.
I did not try to conflate anything, I'm just stating a fact - Reconstructed 4k is 4k, no matter how much you spin it, and you trying to equate my statement to greenburger saying "a 1080p output is the same" is dishonest to say the least.
As far as reconstructing image, the quality and resolution of your source matter a lot more than the image being reconstructed. Here's an example that occurs in video and that you can try yourself
- Grab a RAW 4k video and compress it with a ridiculously high compression ratio - your output will be shit, even though your image will be "real 4k" by your definition, and there's no way around it
- Now grab RAW 1080p of the same video and run it through something like topaz and reconstruct it to 4k - Your output will be much clearer than the "real 4k" video, despite it being "fake" (so to speak) by your definition
- Edit: Your output in RAW 1080p may already be clearer than the compressed 4k, but my point stands.
I thought the same tweet when I read his commentThank you for the explanation Mr. Greenberg, I didn't realize you were still around.
Did I say him/you? No I did not. You guys can't see the forest for the trees here.
When people are posting linked stuff or saying "I heard from someone testing at EA that Jedi Survivor is running at 4K 120hz without a hiccup" without any form of notation that it's not actually 4K it needs to be called out.
I guess "It's running at 4K with DRS and PSSR with Frame Generation" doesn't have the same ring to it...
Last edited:
OP
OP
The same people who were jerking off to DLSS 4k reconstruction are now shitting on PSSR 4k reconstruction.
You love to see it.
You love to see it.
I guess this image as a reply to my comment was a coincidence then.
Thank you for the explanation Mr. Greenberg, I didn't realize you were still around.
Why does it matter that it has DRS / PSSR / DLSS, whatever? a) the presentation is still at 4k and b) nobody gives two fucks as long as image quality is good and performance is stable. I'm sure you have a magnifying glass very close to you to look at those pixels upclose, probably between failed attempts at finding your incredily small penis.
You failed to prove that reconstructed 4k =/= 4k, you then tried to move the goal post and implying that I somehow don't know what image reconstruction is, and now are changing your tune again because there's a load of tools being used to achieve an outcome.
Next thing we'll move our goalpost to "yeah, Horizon Forbidden West may run in 4K at 30fps, but there's Frustum culling so it's not real 4k. Oh, and their fire effects are also at quarter resolution, same as reflections, so guess what, it's not real 4k too!"
So we go from "Game X is running at 4k 120fps" to "Game x is running at 4k 120fps with DRS, PSSR, Frame Generation, Frustum culling, quarter resolution reflections, SOHC VTEC, etc".
You're just digging a bigger hole every time, please continue.
DLSS Quality (1440p to 4k) is fine.
DLSS Performance (1080p to 4k) not.
I still prefer native with good AA implementation but I know DLSS saves a lot of resources... even more in the good AA implementation because most devs uses crap AA (due performance).
First of all, I think DF has it wrong where they mention some custom Sony block for Machine Learning.
The only thing they have in the document is "300 TOPS ML Performance". This is simply regular RDNA3 matrix processing using INT4 variables at 8x the speed of FP32. FP16 = 2xFP32, INT8 = 4xFP32; INT4 = 8xFP32.
In this case, if the PS5 Pro's GPU is running at 2450MHz, we're getting:
60 CUs x 64 ALUs-per-CU x 2 double-ALUs x 2 FP32 MAD OPs-per-ALU x 8 INT4_OPs-per-FP32_OP x 2450MHz = 301 TOPs.
They also mention ML performance as being capabilities of the GPU.
Alex as usual loves shitting on AMD, but everything points to the ML hardware capabilities being simply what AMD already has to offer on RDNA3.
Then he wonders what is responsible for the raytracing performance upgrade, when it's clear in Tom Henderson's leak that the new RT units process BVH8, i.e. at the very least they do twice the BVH intersections per CU.
Also, at ~51m30s in the video:
Guys for the love of god, never ask Alex Battaglia about PC component upgrades.
And this isn't even a game that pushes above 8GB VRAM at 4K maxed out.
People just need to stop using max theoretical TFLOPs as gaming performance metrics. If this wasn't obvious to everyone before we saw how the Series X doesn't run anywhere near 20% faster than the PS5, then it should be now.
This mistake isn't new, though. >20 years ago we were all counting pixel and texel fillrate as metrics to compare console performance. And although they're still very important now (probably the biggest reason why the PS5 matches the Series X in real-world performance is its advantage in fillrate and triangle setup), no one is looking at those numbers nowadays.
The only thing they have in the document is "300 TOPS ML Performance". This is simply regular RDNA3 matrix processing using INT4 variables at 8x the speed of FP32. FP16 = 2xFP32, INT8 = 4xFP32; INT4 = 8xFP32.
In this case, if the PS5 Pro's GPU is running at 2450MHz, we're getting:
60 CUs x 64 ALUs-per-CU x 2 double-ALUs x 2 FP32 MAD OPs-per-ALU x 8 INT4_OPs-per-FP32_OP x 2450MHz = 301 TOPs.
They also mention ML performance as being capabilities of the GPU.
Alex as usual loves shitting on AMD, but everything points to the ML hardware capabilities being simply what AMD already has to offer on RDNA3.
Then he wonders what is responsible for the raytracing performance upgrade, when it's clear in Tom Henderson's leak that the new RT units process BVH8, i.e. at the very least they do twice the BVH intersections per CU.
Also, at ~51m30s in the video:
Guys for the love of god, never ask Alex Battaglia about PC component upgrades.
And this isn't even a game that pushes above 8GB VRAM at 4K maxed out.
It's 33.5 TFLOPs at 2.18GHz.
People just need to stop using max theoretical TFLOPs as gaming performance metrics. If this wasn't obvious to everyone before we saw how the Series X doesn't run anywhere near 20% faster than the PS5, then it should be now.
This mistake isn't new, though. >20 years ago we were all counting pixel and texel fillrate as metrics to compare console performance. And although they're still very important now (probably the biggest reason why the PS5 matches the Series X in real-world performance is its advantage in fillrate and triangle setup), no one is looking at those numbers nowadays.
I don't think it will too.
There is no CPU for that... the same reason for UE Namite e Lumen not running in 60fps too.