they do, the problem is using theoretical peak numbers, they might not translate to the real world use number because of other bottlenecks
Why do you say this?
Good post. :thumbs up:a flop is nothing more than a single mathematical operation, how many mathematical operations it can do on non-whole numbers per second. Processors have more operations than just rote mathematics they can do on non-whole numbers, usually combined into complex operations. These complex operations can use more or less floating point operations to accomplish a similar task depending on architecture, compiler optimization, and task at hand. Its much the same way one can rewrite a single math problem in multiple ways. Like how:
(5*2) + (5*2) + (5*2) = 30 is 5 operations, while (5*2)*3 = 30 is the exact same math, but only 3 operations. Say this was part of a larger calculation on two processors, and processor A does things the former way, and Processor B does things the latter way. And while Processor A can perform more FLOPs, the complex operations it is doing takes more FLOPS to accomplish for the same task.
Really, FLOPS are a very, very poor metric for performance. I've compared it before, but it's like comparing the "power" of two authors in terms of number of pen strokes per second they can put on paper. Author A can put 10 times more pen strokes per second than Author B, does that mean Author A can write better than author B? What does "better" even mean in this instance? In this example, we don't even account for language being written, say Author B writes in Hanji/Kanji and thus every single character takes like 8-9 strokes each, while author B writes in english in cursive, and thus each individual word only takes 1 pen stroke each. Who can write more in that instance, despite author A having more pen-strokes per second?
Things to keep in mind:
A) Floating point math isn't only done in the GPU, CPUs crunch floating point math just as often
B) Ultimately, it's the program written that is using the floating point math. There are ways, for example, to avoid floating point math all together in circumstances.
C) there is also non floating point math which can have different cycles per operations.
Other things to keep in mind: the metrics put forward are FLOPS under the most ideal of circumstance, i.e. every register filled to maximize SIMD performance. There are modes in many processors, for example, which will reserve registers not always available for different tasks. There are always trade offs. There might be instances, like to avoid a cache-miss, where intentionally using every register available isn't very wise.
Because if you compare a couple of PC GPUs with each other you'll realize that it's a worthless performance metric for games. It's why Nvidia & AMD never really advertised it either before this gen.
I think part of the reason it became popular is because it's an easy to understand thing for non technical people, which most console users are. Bigger is better is a very simple concept. Too bad it just doesn't work that way.
I'm not saying anything about the link between teraflops and sales numbers. Sony said it was something important to consider when buying a PS4. The higher number meant better graphics but now it doesn't mean better graphics.Don't forget the second half of last gen the Xbox X was the most powerful console and Sony had no problem continuing to dominate sales. The first xbox was stronger then the ps2 and 3rd party games were better on the 360. Power don't decide shit.
Care to explain why?
That's what Jason Schreier was saying and he is tech minded and knows a lot of devs.
Both CPUs are roughly a Ryzen 3700X that doesn't boost as high. The cache setup may be different, but this hasn't been confirmed yet. The PS5's GPU is a higher clocked RX 5700 and the XSX doesn't really have a desktop counterpart yet, but it's more or less 30% faster than a 5700XT. Both GPUs do have newer features than current Navi/RDNA.can someone break down the following in terms of current on the shelf PC Parts. So this TFlops thing makes more sense to me.
PS5 = *this GPU/CPU (roughly)
XSX = *this GPU/CPU (roughly)
So why aren't these the same as MHz / GHz? (i.e. the number of processor cycles per second)? Simply because cycles do not correlate 1 with floating point operators; different processors take more or less cycles to perform one FLOP (in fact, some processors take several cycles for a single FLOP, while others can perform several FLOPs in one cycle). In the end, FLOPs are just a better indication of power, especially in the context of 3D graphics, because that's what you want the processor to do; you don't care how many cycles it takes.
Like any complex systems, bottlenecks happen. High TF won't mean jack if the rest of the system is slow.
Disc speeds, drive speeds, these things matter too.
This sounds like a stealth Sony defense post, but I have no idea who's got what FLOPS, or faster drive or anything. Just pointing out what Cerny said has merits.
A lot of people missed this. And I think this is an important distinction
this Is very true and the entire ecosystem tells a bigger story than just the GPU alone. I'm not ready to crown which system performs better.
Someone else talked about the bottle neck of data transfer. So that is a game changer on top of the GPUs and again someone else spoke on that being one general task. Whereas one console may be better than the other depending on the specific task.
Also I believe that at 4k 30 and maybe even 4k60 they perform exactly the same and we may only see the differences when they are pushing 8k.
honestly without seeing games first party and/or multiplat we aren't going to really know who does what better. On top of how easy it is for developers to tap into the features of each architecture brings to the table.
this is kind why the console warz are dumb. End of the day it's the games that are created and how they are utilized on the platforms that tells the whole story.
exactly this. like even while both have the same base architecture there are many tweaks taken, some different efficiencies made. even funny enough if you look at the XSX memory speeds and took an average of the two the average of the two they equal the 448 G/s that sony has ( I do know there are probably performance gains with the higher speed 10 gigs vs the lower speed 6 gig but again it all matters how that is utilized) s this is why like while we have all these specs, the real comparison will be based on what devs do at the end of the day. Im excited to see just how both approaches work and where they lag in gave development, but really the introduction of an SSD on both platforms I think will be a great thing for all games going forward and all the efficiencies brought on because of that.Oh yeah, absolutely. (Tera)FLOPS are better than clock speed as a metric, but are still just one metric in a sea of many, many other inter-related elements that govern total performance. It's next to impossible which of two (closely-specced) machines will perform better unless you're incredibly knowledgeable about all aspects of computer architecture... or you simply test their performance against each other Digital Foundry style, which obviously we won't be able to do until they release.
The bottom line is: if performance of non-exclusive games is an important factor of your purchasing decision, you have no choice but to wait until both of them are released and benchmarked. And even then it's likely to vary on a game-by-game basis.
I fee like this should be added to the OP. To take the writer analogy further, I'd conclude that you could probably say that Author A is better than Author B if Author A can write 10 pen strokes per second vs .05 pen strokes per second. Author B might produce some great works, but it'll take forever for them to create anything. That's how I view the difference in the current gen vs next gen.a flop is nothing more than a single mathematical operation, how many mathematical operations it can do on non-whole numbers per second. Processors have more operations than just rote mathematics they can do on non-whole numbers, usually combined into complex operations. These complex operations can use more or less floating point operations to accomplish a similar task depending on architecture, compiler optimization, and task at hand. Its much the same way one can rewrite a single math problem in multiple ways. Like how:
(5*2) + (5*2) + (5*2) = 30 is 5 operations, while (5*2)*3 = 30 is the exact same math, but only 3 operations. Say this was part of a larger calculation on two processors, and processor A does things the former way, and Processor B does things the latter way. And while Processor A can perform more FLOPs, the complex operations it is doing takes more FLOPS to accomplish for the same task.
Really, FLOPS are a very, very poor metric for performance. I've compared it before, but it's like comparing the "power" of two authors in terms of number of pen strokes per second they can put on paper. Author A can put 10 times more pen strokes per second than Author B, does that mean Author A can write better than author B? What does "better" even mean in this instance? In this example, we don't even account for language being written, say Author B writes in Hanji/Kanji and thus every single character takes like 8-9 strokes each, while author B writes in english in cursive, and thus each individual word only takes 1 pen stroke each. Who can write more in that instance, despite author A having more pen-strokes per second?
Things to keep in mind:
A) Floating point math isn't only done in the GPU, CPUs crunch floating point math just as often
B) Ultimately, it's the program written that is using the floating point math. There are ways, for example, to avoid floating point math all together in circumstances.
C) there is also non floating point math which can have different cycles per operations.
Other things to keep in mind: the metrics put forward are FLOPS under the most ideal of circumstance, i.e. every register filled to maximize SIMD performance. There are modes in many processors, for example, which will reserve registers not always available for different tasks. There are always trade offs. There might be instances, like to avoid a cache-miss, where intentionally using every register available isn't very wise.
you know a thread is derailed when even krejlooc post memesteraflops don't matter if all you do is 16.16 fixed-point math