Haven’t you heard, the SSD is so fast, you don’t even need RAM!
1 Like
The amount of mental gymnastics Playstation fanboys are doing to defend their precious piece of plastic is staggering to say the least. While the PS5 is RDNA 2 based its a bit of a mix of 1 and 2 seeing its missing some RNDA 2 hardware features. People arguing over the PS5 , RDNA 1 and 2 are missing the point entirely. Basically these so called Sony insiders or Tech youtubers who come up with this info have never come close to being right about anything to do with the PS5. This misinformation could have easily put to bed months ago if Sony had released an official die shot and a list of supported hardware features instead Sony let its rabid fanboys argue info supplied from fake insiders.
1 Like
Bottom line the Series X is more powerful and has a more advanced GPU, Xbox fans do not have to go through the start of another gen where the Xbox console is under powered and mocked by meme’s. Because of how Sony fans dog piled at the start of last gen it seems they are trying to build a wall of lies so that do not have to deal with the mocking they put out at the start of last gen.
The Goal posts keep moving and moving, first before Spec reveal the PS5 was a 13 TF monster and would crush series X, then after spec reveal TF’s did not matter and it was all about the SSD and Tempest. Turns out their audio solution sucks and was only developed to get around Atmos license fee’s and the SSD speeds are not much better then Series X load times.
Then for awhile now it was RDNA 3 would save the day and the PS5 had hidden features not revealed yet. Now after the die shot it turns out the Series X SOC is more advanced and actually has more future thinking features like VRS, SFS, and Hardware built for ML/Super sampling to counter DLSS. It does not look like the PS5 will have any hardware assisted solution for a DLSS like feature. Series X is more powerful and more advanced and that seems to REALLY bother the Sony crowd that wanted another easy slam dunk at the start of this gen.
7 Likes
I remember some of the insanely hyperbolic things he was saying after Cerny’s initial talk and now I’m sitting here thinking, what on earth was he talking about??? Ridiculous.
1 Like
Yes, and those who pointed out his fanboism was accused of being console warriors… interesting how things went then. 
3 Likes
It’s interesting that Mark Cerny was out and about after the PS4 launch giving detailed technical break downs at chip level in interviews Link but this time out we’ve had nothing on the CPU or GPU at this level of detail.
That alone tells it’s own story. I certainly was never convinced there was anything in all the IC and RDNA3 features talk. The sources are certainly dubious, just take a listen to the MLID podcast where NXG guests for some context into that channel.
3 Likes
Amazing indeed.
It was like it was hard to say that the Xbox Series X was clearly the more powerful machine.
The other place has those people also, with “dev” posters continuously highlighting that the power delta is less that it has ever been between two consoles. Many stated that there would be no tangible difference between the two. It was funny how another dev chimed in stating that the extra horsepower would be visible…
Here’s a question: If Penello still does not get a pass for things he said 9 year ago, then why are we listening to these chaps?
3 Likes
Sony was ambitious with their tech design up until PS4, but PS4 itself was not all that ambitious. It was largely PC parts with a few tweaks. It’s upper hand from a tech pov mostly came down to blind luck wrt RAM supply finding a workable price point last minute. I dunno how Cerny got his mythical storyline that Sony fans attach to his capabilities to design hw. Pro was not real well designed for its purposes and fell short in a lot of ways. PS5 looks very similar to that.
Anyone stating that is foolish or deliberately tryna smooth over very real differences. It is objectively not true. Even if you compare the delta to X1X/Pro it still isn’t true.
2 Likes
Yea Cerny has been out engineered by MS twice in a row now and the only reason the PS4 was so much better was he lucked out on ram prices at the last minute. Had the PS4 had half its ram the Xbox One would have been more on par.
It’s just typical for a Sony fan to hold a Sony exec in a higher place then they deserve, you’ve seen it for a long time with all the Kaz, Ken gifs and memes,
1 Like
Don’t get to excited about DirectML the Xbox is never going to come close to what DLSS is able to do. The Xbox doesn’t use AI teacher data like DLSS uses, its using 4 and 8 bit integer for its ML. Another thing DLSS uses hardware acceleration through Tensor cores on nvidia GPU’s but I don’t know where hardware acceleration is happening on Series X/S.
1 Like
I think you’re confused. DirectML by it’s very nature uses “AI teacher data”, machine learning is literally teaching a computer how to do something. How do you think Auto-HDR works? They’ve had to teach the API how to apply HDR.
Don’t think they’ve really given much detail, we know it’s there as it was mentioned at hot chips as a small area cost to give 3-10x performance uplift. But no details.
I’m no expert but from what I’ve read the Xbox uses Inference AI where is Nvidia’s DLSS uses teacher data combined with a Artificial neural network to learn.
Not sure what your getting at here, I guess you are trying to say that the ML HW in the series x is not as strong as that found in the Nvidia cards (and this may be true). But I don’t see what that has to do with how the software operates. DirectML is HW agnostic if your game runs directML it will run on anything Nvidea cards, ADM ones etc. Just at different performance levels. The directML demo was running on a titan V but that same code would also run on AMD 7000 series + or xbox series x or a potato laptop with integrated gfx card…
1 Like
The benifit of DLSS is that from turing onwards the Nvidia GPUs have hardware seperate from the rasterisation hardware that does the DLSS math. RDNA2 gpu’s share the burden with rasterisation. So Ai reconstruction will come at a greater cost compared to DLSS.
However its still could produce significant benefits, for example say id software are making a new doom they could target a 1080p internal resolution which uses 8tflops of the GPU and the use the remaining 4tflops to reconstruct to 4k.
I dont know which method would give more resources for geometry and effects, but I think directML will offer more performance for more geometry+effects then just aiming for a higher resolution.
It would be even more crazy if they could combine image reconstruction with dynamic resolution scaling + VRS, with a combo like this it would be like a 40tflop console.
1 Like
When comparing to DLSS i believe there are 3 points to consider to compare performance with Xbox ML super resolution.
-
Efficiency of the network - No one can tell here whose network will be better at predicting super sampling. It’s the biggest variable.
-
TOPS - fixed value of operations which can be performed by the hardware. Nvidia’s DLSS is based on int 8 and series X can also do int 8. Here, series X is comparable to 3060 i suppose.
-
Parallel work of AI hardware - This is where Nvidia DLSS has clear edge. It’s tensor cores works in parallel with shaders core. Same is not possible on series X because it uses shader cores as AI cores. Although, super sampling works at the last point of the frame time so after all the work is complete from the shader cores. Hence, if parallel nature matters or not is unknown.
2 Likes
AI has 2 phases: learning a neural network model and using it. The second part is called Inference. DLSS ingame is Inference.
4 Likes
I think it’s too soon to say anything regarding this.
It could be as you said and hardware resources may have to be split to use Super Resolution
Or
It could be just be a penalty on frame time and the full resources could be available for both rendering and super resolution.
Its never to soon for speculation 
One thing youre missing comparing to Nvidia is with Nvidia the number of Tensor cores are fixed and generally, thinking from memory, number from 100-300 cores as a rough and varying estimate depending on model and series…whereas, although the same given GPU cores on the XSX can’t do DirectML and rasterization in parallel…they can however set aside any number from the 3,328 shader cores and possibly leverage any amount they want , balancing that leveraging for DirectML with leaving the majority for rasterization. Point being, hypothetically they could leverage say 1,000 shader cores for DirectML, render at a lower resolution with the remaining majority, and then up res from there. Different approach. The cores for DirectML can be selected and balanced with rasterization out of the entire pool. Though obviously 1 tensor core is not equivalent to 1 RDNA2 core but just wanted to point out the computational DirectML power is not set but can be balanced out amongst a number of configurations.