Simply embarrassing. This is what happens when you don’t read the updated op.....
ThisWe all knew it was custom based RDNA. Basically its RDNA5 because PS5.
I read the update and while it has information from the Sony developer clarifying what he meant, I can't find anything in the update that states the information contained in the original OP was faked, photoshopped or a product of the shadowy Xbox Discord.
What's embarrassing is the behavior of a certain group of fans that when confronted with information they don't like they attack, accuse, threaten, report and invent conspiracy theories.
This is also achiveable with lower power. 300W is only needed if the chip is stressed. As far as I understand, the PS5 chip will clock up if it is not fully stressed (else it would really suck the power out of the wall). Also they seem to have cut out a few things. This makes the chip less complex and it can take more power. E.g. MS added (or let AMD add) support for Int8, Int4, ... instructions. This also makes the chip more complex which will result in worse clock-rates (than without such things).It's achievable. With ~300W power consumption (5700xt). I don't think that they put something that hot into PS5, but how much were they able to cut it down? To have some normal thermals they would need to cut that power consumption by 1/4 at least (Is that even possible?).
But but but it has a fast SSD
But also like @Bill O'Rights said the way they posted the information was pretty bad. They should have had it verified and with proper direct links before posting it. It's normal that some people were going to question.
It's like the leak that Nikana gave us. Nobody questioned it because he had it vetted. If the information had direct links and the DM was verified by the mods nobody would question it.
I read the update and while it has information from the Sony developer clarifying what he meant, I can't find anything in the update that states the information contained in the original OP was faked, photoshopped or a product of the shadowy Xbox Discord.
What's embarrassing is the behavior of a certain group of fans that when confronted with information they don't like they attack, accuse, threaten, report and invent conspiracy theories.
To be fair me, quite a few people still questioned it but didn't do so with the same veracity. They didnt question the information came from a source that was vetted but they did question if it was true or not since there is an NDA and other tried to convince me that I made a mistake posting the information. Some of the very same people who cried fake in this incident.
That Sony engineer has probably been forced to give a statement to a media outlet by HR and got into trouble because of delusional warriors who cant just keep it to the forum.
Hopefully that isn't true because I would hate for anyone to get in trouble over this.
So what RDNA 2 feature is missing?
Well whenever information that is posted breaches an NDA I think it's normal for some people to be concerned. I definitely was a bit rough on O'Dium because I thought he was getting someone in trouble by posting that DM.
I love getting inside information but not at the cost of somebody's job.
MS customized the XSX HW to support 4bit, 8bit and 16bit integer rapid packed math just for Machine learning and the XSX also supports DXML. Likely not as good as Tensor cores but an in between of sorts.
Also Cerny has already stated that PS5 supports Primitive Shaders.
Basically the predecessor to Mesh shaders.
Dictator said:yeah about half of the ML performance of a 2060. So DLSS on an XSX would cost around 5.5 to 6 miliseconds of compute time if it scaled like it does across Turing to RNDA 2.
Big enough to say he was wrong and apologize, I tip my hat to you, good sir, it was wise to take the precautions and I hope as I'm sure most do that it all ends well.
Yes it obvious you don't know where to startThis is so bad, I don't know where to start...
there is no maybe"DLSS 2.0 runs on the (tensor/ML cores) " - maybe it does, maybe not.
"Also DLSS 1.0 runs on the shaders of the GPUs and performs noticeably worse." - it "performing worse" (as in looking like shit, compared to plain FidelityFX, which, by the way, is crossplatform) has nothing to
It seems that you are aware that most training against large datasets is done in the cloud but you're unaware how the inference part is executed on hardware.do with exactly how neural network inference was executed.
"hardware accelerate ML good, "software-based ML" bad." - what the heck is "software based ML" on GPUs, clueless one?
Make of that what you will.
Make of that what you will.
100% this guy will be having a meeting with their manager TODAY over this. Sony like most software companies have a strict "no talking publicly" about products unless PR trained and authorized. So while the dude did not intend to violate this, he was dumb and we saw the result. I am sure his team is looking at him like "WOW, maybe you are not as smart as I thought you were"
Not even a counter-argument and is absolutely meaningless within the context of what we are discussing.Welp, make of NVidias 2015 claim that GPUs excel at neural network inference what you will.
Both PS5 and Series X use custom RDNA2. They both are basically RDNA2 but will have some extra things and some removed things.So we could end up with a situation where we have one console with 10.2 RNDA1.5 tflop (variable) vs one console with 12.2 RDNA2.0 tflop (sustained)?
Perhaps the gap is bigger than the 17% people talk about.
OR its the same for Xbox Series X - it also use RDNA1.5?
There is all maybes in the world.there is no maybe
NVIDIA DLSS 2.0: A Big Leap In AI Rendering
Through the power of AI and GeForce RTX Tensor Cores, NVIDIA DLSS 2.0 enables a new level of performance and visuals for your games - available now in MechWarrior 5: Mercenaries and coming this week to Control.www.nvidia.com
Well, that is hardly surprising.[I have severe reading comprehension problems]
No, not "most". ALL.It seems that you are aware that most training against large datasets is done
It likely doesn't mean what you think it means.but you're unaware how the inference part is executed on hardware.
I think what bothers me the most about this was that it was shared without his consent. Unlike Nikana leak where he got permission to share that information. The developer in Nikana case agreed to take a risk by releasing that information. I'm pretty sure Rosario never wanted that information to be shared which is why he DMd O'Dium instead of sharing it publicly via a tweet.
This thread be spawning fanbois from another dimension like
I don't think you understand that TENSOR cores are fasterWelp, make of NVidias 2015 claim that GPUs excel at neural network inference what you will.
Please educate yourself on the comparative performance of Nvidia's Tensor Cores vs MS's solution, as you seem to be overly flattering with your claims about MS performance in this area:
From Ree.
Also, Series X doesnt have specific hardware to accelerate it, basically like DLSS 1.9 the ML calculations are done using the shader cores. RTX cards dedicate a significant portion of their GPU die for the tensor cores.
Yeah, "mabye it runs, maybe not" is so hard to comprehend, I mean, it takes a genius to translate "NN inference rusn great on GPUs" into "as it could run on normal SMs as well".Not even a counter-argument and is absolutely meaningless within the context of what we are discussing.
I'm not only fucking understand that, I have literally REFERENCED exactly HOW MUCH FASTER they are IN CERTAIN TASKS in my FUCKING PREVIOUS POST.I don't think you understand that TENSOR cores are faster
Nothing you stated is contrary to what I stated.
We agree that Tensor cores are the superior method of using HW accelerated ML.
Not sure why you're asking me to educate myself.
I stated its an in between fully custom HW block accelerated like tensor cores vs just using Shaders to do the calcs.
Yea its using the shaders but at a much reduced cost when you're using 4bit, 8bit and 16bit RPM.
That's why I stated an in between.
Like I said the only thing in contention here is if DLSS2.0 runs on a special accelerated part of the GPU [tensor core] (it does).There is all maybes in the world.
NVs claim it has to be tensor cores (which is nothing but basic 4x4 matrix multiply-add, something that GPUs, as NV itself figured in 2015, excel at) are as valuable as raising price on TEsla cards now, because "mining".
They have no other way to justify not having it on pre-Tesla GPUs.
[edit:Let's not be so harsh on you.], the link you posted before is local benchmarks inferencing, you can train and or run inference locally on a machine with TensorFlow, PyTorch or whatever you chose to use.Well, that is hardly surprising.
No, not "most". ALL.
FFS It easy to spot the 101's good luck with your course.It likely doesn't mean what you think it means.
Actual ML benchmarks with Tensors on vs off (so shaders show that using tensor cores is mostly, although not always, faster, but even when faster, it's about 2-3.5 times, not tens of times.
There is no magic on having neural network inference run on non-tensor (which sound cool, but is actually dumb, superstraighforward packed math) on cards without tensor cores, certainly not for the faster cards, such as 1080Ti.
You made this up, there is absolutely no source on this planet who claims that DLSS 1 runs on "shaders", while 2.0 on TCs.DLSS1.x runs on the shaders
You can train or run inference locally on pretty much any computing device, including my dish-washing machine.you can train and or run inference locally on a machine with TensorFlow, PyTorch or whatever you chose to use.
They've called it 'packed math" and it started with Vega, I think.- added instructions for 4-bit, 8-bit, 16-bit dot products (my guess SIMD instructions specific for ML)
Hard to say by how much, but they reduced it by quite a bit. The XSX power supply is 315W, and they are rocking 1.8GHz and 52CUs for the GPU. According to AMD they achieved a 33% decrease in power consumption for the same performance compared to RDNA1.
That was about Radeon 5700XT and in response to info, that it's not able to reach that clock speeds.This is also achiveable with lower power. 300W is only needed if the chip is stressed. As far as I understand, the PS5 chip will clock up if it is not fully stressed (else it would really suck the power out of the wall). Also they seem to have cut out a few things. This makes the chip less complex and it can take more power. E.g. MS added (or let AMD add) support for Int8, Int4, ... instructions. This also makes the chip more complex which will result in worse clock-rates (than without such things).
What really struck me is the sheer size of the PS5 case. This could mean they have a big power supply and a big cooling solution. The other parts aren't that big. I assume the PS5 also contains a power supply in the range of 300W. Else they wouldn't have such a big case.
But that's not what he said. He said more features with one cut.If this is indeed accurate, it seems to br RDNA2 with cut features, effectively making it not full RDNA2, hence "between RDNA and RDNA 2".
What he said wasIf this is indeed accurate, it seems to br RDNA2 with cut features, effectively making it not full RDNA2, hence "between RDNA and RDNA 2".
You’re late to the thread.But that's not what he said. He said more features with one cut.
What he said was
"It is based on RDNA 2, but it has more features and, it seems to me, one less."
I questioned the ethics of chasing a sony engineer on social media, getting a Personal DM saying he needs to check if its OK to talk about it due to NDA, and then posting it on GAF.
That Sony engineer has probably been forced to give a statement to a media outlet by HR and got into trouble because of delusional warriors who cant just keep it to the forum.
Not Classy. Some people need a reality check to mess with others livelihoods on social media.
Because it's not in your head does not mean it does not exist. time and space exist outside of you.You made this up, there is absolutely no source on this planet who claims that DLSS 1 runs on "shaders", while 2.0 on TCs.
With a target resolution of 4K, DLSS 1.9 in Control is impressive. More so when you consider this is an approximation of the full technology running on the shader cores.
I said :You can train or run inference locally on pretty much any computing device, including my dish-washing machine.
Training the neural netowrk in the context of our discussion, namely rendering games at higher resolution, lower resolution and using them as input/output to determine the NN weights is not something even remotely reasonable to be done on a gaming machine. It is completely done in data centers by NV.
It seems that you are aware that most training against large datasets is done in the cloud.
which is proof you're clueless.No, not "most". ALL.
Man I would think you are different person from the one who was responding to O'dium on his post about his explanation and then later on his discussion with this engineer. You came way insulting and condensing. In fact I would say you were the catalyst that caused that thread to turn into shit show and him having death threats to him and his wife.
Or do you think this was justified because you believe he did something wrong?
Not sure if I buy this. I worked in the media sector of gaming for many years.
People dont just talk when asked about things they aren't sure if they can or can not talk about. Even the developer I talked to, it was laid out very clearly what I could and could not say. We talked about more than I posted but it was very clear what I was and was not allowed to say.
I have known this person a long time and they know I would never just blurt out information which is why they were willing to talk to me. The scenario in question with odium would have to have a lot of coincidences for it to be a screw up. Unless Odium/Gavin was purposefully willing to burn that bridge for the sake of a Neogaf post, there is no way I buy that the information was given with this common midnset that it would not be shared.
Because it's not in your head does not mean it does not exist.
Compared to CU/SMs, which are programmable, it's actually rather straightforward.Tensor cores can do a huge 4x4x4 (my guess 4x4 with accumulate) calculation. That's a lot of transistors to pull that off.
My guess, that Nvidia's solution runs concurrent with the shader cores, and so rendering is not impacted in terms of processing, but bandwidth is (as competition for that resource).
Man I would think you are different person from the one who was responding to O'dium on his post about his explanation and then later on his discussion with this engineer. You came way insulting and condensing. In fact I would say you were the catalyst that caused that thread to turn into shit show and him having death threats to him and his wife.
Or do you think this was justified because you believe he did something wrong?
Heads were rolling when people speculated ...
He could have just made the information public instead of tweeting it. That's what he did with everything else that he shared with O'Dium. I mean if he didn't have any issues with the information being public why would he feel bad about what he said?
How do we as fans of these machines lose out at that?
We don't and thts the point. The fud narrative of one system not having certain features when this same narrative was also pushed this gen with dx12 and Sony smashing tht frivolous agenda to oblivion is my point. When I mention history you do know in what context I'm saying it right? As in propoganda like this has been spread before and Sony came out with similar or better solutions, apis, features, etc. First it was Sony doesn't hace RDNA 2 thts been debunked, no ray tracing lol tht was annihilated with almost every game shown having it during presentation and list goes on. In name Sony might not have certain important features buttttt be sure thy will have important similar or better features, thy have proven this in the past. Just because one company may do something better tht isn't the end of the world. Tech isnt always equal, if it was it would be a boring boring world.
I read the update and while it has information from the Sony developer clarifying what he meant, I can't find anything in the update that states the information contained in the original OP was faked, photoshopped or a product of the shadowy Xbox Discord.
What's embarrassing is the behavior of a certain group of fans that when confronted with information they don't like they attack, accuse, threaten, report and invent conspiracy theories.
The fact is we don't have conclusive evidence to suggest either XSX or PS5 is the full RDNA2 part/standard. They're both custom solutions based on RDNA2. To pick and choose which rumor/fud to agree with is embarrassing. To jump to conclusion is also embarrassing.
Hello. Thanks for the friendly and cordial discourse right now compared to the rest of the console boys going off their heads around us.
I can kinda understand why you'd think that I'm pushing a narrative but all I am saying is that both have taken the same basic building blocks of the engine and chassis and have 2 different design principles around it to do the job required.
The MS route is stated as highest and most powerful so everything else had to be designed around that. But that requires innovation too. The fact that CPUs, GPUs and everything that makes them improve performance is because of innovation. RNDA 2 is an innovation. Zen 2 is an innovation.
Anything AMD, Intel or Nvidia do to improve on their CPUs and GPUs is an innovation. I get that innovation wise MS have been working on VXA, SFS and some audio technique. But they mostly seem to be software based.
Now PS are saying we're not going for top of the line power so all this extra engine room and chassis space can be used elsewhere on the car. So their APU has inbuilt controllers and processers to do stuff that other stuff that helps overall balance.
I'm just looking it at an engineering POV. It's like back in the 90s we had separate audio cards until everything was offloaded onto the CPU and Audiocards became niche. But Sony are saying with the Tempest engine is that "Hey, remember when sound was processed elsewhere before instead of the CPU? Well we've put sound back on a dedicated chip, so that means free CPU cycles for other processes."
Which from the engineering perspective is cool because it's about the balance. The PS5 is less powerful SOC overall at top end but it is supplemented by other techniques.
So it's more hardware based.
I'm old enough to been through 5-6 console generations and the stupid wars. I know it's about the games and enjoyment and any console war is fraught because it's not easy to own 2 machines so praising your purchase and putting other people's choices down is a self image thing that I don't care about. For me it's about the tech and how gsmes developers will leverage these new developments. But as someone interested in the tech, I can't help but be interested in the entire fact of Mark Cerny's design is "SEE HOW WE'VE ELIMINATED ALL THESE DESIGN BOTTLENECKS!"
And I do want to see...