The article says it's time sliced, so it would get 100% of the GPU for 10% of the time.
Is it supposed to be fixed or dynamic?
Digital Foundry Article Technical Discussion Archive [2013]
- Thread starter helio
- Start date
-
- Tags
- digital foundry
- Status
Is it supposed to be fixed or dynamic?
There can be multiple ways to do this, however it might be something like the time slice on the CPUs for the background GPU command issue thread or threads is roughly ten percent of the run time.
Maybe the GPU itself can internally time slice, but it seems disruptive to think that the GPU as a whole dumps everything it is doing before and after the time slice.
GPUs traditionally do not context switch well or permit interruptions of running shaders--although the term has popped up in a limited case in some CI documentation for a CU-level flag for resuming a compute wavefront.
If the kernels can't be swapped out, then waiting for them to complete may not happen in the time window.
Preemption is an eventual goal for AMD, although it's not listed for this gen.
Getting some kind of clean slate might eat up a significant chunk of the time slice and disrupt the performance of the game slice by forcing it to spool up again after every transition, and the benefit over just letting the automatically load-balancing GPU hardware do what is necessary seems minor.
The catch to the latter case is that getting a good guarantee of prompt scheduling would require some kind of allocation of graphics and CU resources being held in reserve so that they can be pulled in as quickly as possible.
That is something I've discussed elsewhere.
Maybe the GPU itself can internally time slice, but it seems disruptive to think that the GPU as a whole dumps everything it is doing before and after the time slice.
GPUs traditionally do not context switch well or permit interruptions of running shaders--although the term has popped up in a limited case in some CI documentation for a CU-level flag for resuming a compute wavefront.
If the kernels can't be swapped out, then waiting for them to complete may not happen in the time window.
Preemption is an eventual goal for AMD, although it's not listed for this gen.
Getting some kind of clean slate might eat up a significant chunk of the time slice and disrupt the performance of the game slice by forcing it to spool up again after every transition, and the benefit over just letting the automatically load-balancing GPU hardware do what is necessary seems minor.
The catch to the latter case is that getting a good guarantee of prompt scheduling would require some kind of allocation of graphics and CU resources being held in reserve so that they can be pulled in as quickly as possible.
That is something I've discussed elsewhere.
ProjectNatalFan
Newcomer
simply put
Is what goossen says impossible or possible? Or need more info. Nothing is ever cleancut it seems lol
Is what goossen says impossible or possible? Or need more info. Nothing is ever cleancut it seems lol
There's nothing outlandish or impossible about it.
They are free to design the system with a reserve, and they are free to design said system with the ability to tweak it up or down.
It's not clear how many knobs there are for the GPU's arbitration policies, but they can be very involved.
The time slicing for background threads and their priority with the GPU can be modified.
Resource reservations for storage and pipeline allocations can be tweaked.
I don't think anyone really thought much about it until the DF article came out about it. MS decided to discuss it when I'm not sure anyone was asking about it. Perhaps DF did?
It was 10% on the old 360 too?
Sony didn't say, but MS did. I think that's why one is known to us.
What if it's a very coarse slicing, maybe a full 1.6ms at every vblank, the harm done for wiping it clean would be minimal because it would be when the game code is restarting a new frame anyway. And it would help kinect get a very low latency, the game code would receive kinect data right at the moment it needs it to start a new frame.
That's how it was done on the 360. I have no idea how it is done on the X1.
ProjectNatalFan
Newcomer
Thank you. I guess it is possible which is good news. I dont think they would come out and say this if it was impossible. maybe we get more info about this in the article over the weekend. Another question i would have does he imply the whole 10 percent or a portion.
Compute kernels could be running out of step with rendering. One thing of interest is that some documentation on Sea Islands that has been posted indicates there is a information indicating when a compute kernel needs to start back up again after a context switch.
This could be related.
Also, if the rendering engine's target rate isn't 60 Hz, does the slice grow to keep the percentage the same?
If the game starts chugging, the system's outlay can become similarly variable.
Sony refused to comment on it when DF asked the question.
I'm pretty sure if was Yoshida who said something to the effect of we don't discuss system reservations.
djskribbles
Legend
I think that was just in regards to the memory reserve, after DF published their article (if that's what you're referring to).
djskribbles
Legend
They didn't. Matter of fact, several sources have said that it's more than 5GB available for games. I'm just saying that's the only time I've seen Sony say they don't want to talk about reservation numbers was when DF posted that article.
The irony is that people would choose to believe anonymous sources over the actual person that designed the damn thing ;-)
Sony didn't deny the actual numbers though, while taking time to deny/explain some side issues.
Essentially confirmation imo.
No platform holder ever confirms RAM reservation that I know of. Last gen too.
- Status
Similar threads
