They need SFS to reach the stated figures and therefore don't developers need to implement it at engine level? Don't know if it's as simply as flipping a switch.
SFS is an extension of something already around on last generation and likely a part of any modern game engine. The customisations for XSX augment it. The principal of only loading in parts of a texture file you need to effectively boost IO bandwidth/save RAM isn’t unique to XSX.
Sony would likely have profiling tools that do give them the clock speeds.
From that it's a simple calculation based on the number of CUs to get to TFs, which are always just a theoretical number based on a calculation and not a real measure of the actual calculations being made (as that would require CU utilization numbers).
(not saying the rumor is valid, almost all rumors here are complete BS lol)
It’s entirely bullshit. The only question is whether the guy is unknowledgeable and being taken for a ride by someone else before posting it here, or whether he’s unknowledgeable and trying to take people here for a ride. Either way it’s a #BAMO from me.
Secondly it’s not at all a case of looking at clock speeds and doing some quickmaffs to arrive at some 9.333 (recurring, of course) teraflop number as no game ever even gets near 100% CU occupancy. Synthetic stress test dumb loop code that does could flip each transistor every cycle to technically carry out 10/12 teraflops of useless calculation would cause even the XSX to melt or thermally halt.
Quoting a developer as themselves quoting teraflops figures is
laughably stupid. If for some bizarre reason they wanted to do that, they’d have to work out CU occupancy for a particular frame and then try and estimate it. It’s a metric only crazed fans care about, and that is the real source for this “rumour”.
It’s been said many times, but fixed clock doesn’t mean fixed workload per clock tick. It doesn’t mean fixed “teraflops” of calculations being done per unit of time.
There’s a reason a CPU or GPU can be at a fixed frequency and get hot when running a benchmark and be cool to the touch when idling.
Calculation—or work done—consumes electrical power and produces heat. Power and heat can and does vary based on the code being done, even on fixed clocks.
Targeting fixed peak power consumption is a better philosophy than targeting fixed clocks with variable power consumption.
You can’t measure clock speed to determine how much work is being done. You can’t measure clock speed to say “how many teraflops a console is.. hitting or struggling to hit(?!)”.
You can measure power consumption to get a ballpark idea of how much work is being done. Something you can do on any modern gaming PC. Something you can even feel in the room as heat if you cohabit with a ThreadRipper. A ThreadRipper that when running as close as possible to its peak attainable work rate will actually be at a lesser clock than at a single threaded peak load.
You cannot measure how much work its doing by frequency, but you can see it graphed out plain as day in power consumption on a decent PSU.
If the power consumption is staying pegged as high as the chips TDP can handle then you’re at a fixed and maximum work load, regardless of frequency fluctuations.
Your hypothesised developer would be better off profiling power consumption than frequency. In reality they’ll have a GPU utilisation metric to monitor that accounts for it.
That developers are talking about “hitting” teraflop numbers is utter nonsense. Complete and utter tripe as presented by the idiot here vaguely claiming to be quoting whispers in the wind to lend it credence.
The reality of it is that game code—even efficient code with high CU occupancy—rarely loads the CU above 30-40% per unit of time. The rest of the time they’re waiting on results of other things being completed, or are waiting for a cache miss to be rectified by a memory fetch etc.
Neither PS5 no XSX will ever get near their theoretical synthetic stress test maximums while playing any kind of game that reacts to input and is doing lots of different tasks.
How close they get to actually efficiently using the CUs at around 40% will depend way more on the developers, the different graphics APIs, and the slight differences in RDNA2 customisations and cache/memory architecture than the differences in peak figures.
Something people in the know and developers have been saying since before we knew what we were getting. The same people saying the difference between the two is remarkably close.
Closer than 360 and PS3 were, with even less wildcards I’d imagine as the architecture is so much more similar.
Buy for the games, controllers, friends etc. The only things that actually matter and have any real difference.